#  @MaaSAI42 ModelScope ModelScope posts on X about ai, native, agentic, inference the most. They currently have [-----] followers and [---] posts still getting attention that total [------] engagements in the last [--] hours. ### Engagements: [------] [#](/creator/twitter::1784494412913049600/interactions)  - [--] Week [------] -12% - [--] Month [-------] +72% - [--] Year [---------] +210,270% ### Mentions: [--] [#](/creator/twitter::1784494412913049600/posts_active)  - [--] Year [--] +9,400% ### Followers: [-----] [#](/creator/twitter::1784494412913049600/followers)  - [--] Week [-----] +4.50% - [--] Month [-----] +52% ### CreatorRank: [-------] [#](/creator/twitter::1784494412913049600/influencer_rank)  ### Social Influence **Social category influence** [technology brands](/list/technology-brands) [travel destinations](/list/travel-destinations) [stocks](/list/stocks) [finance](/list/finance) [products](/list/products) [gaming](/list/gaming) **Social topic influence** [ai](/topic/ai), [native](/topic/native) #166, [agentic](/topic/agentic) #244, [inference](/topic/inference) #72, [llm](/topic/llm), [model](/topic/model), [realtime](/topic/realtime), [the first](/topic/the-first), [math](/topic/math), [strong](/topic/strong) **Top accounts mentioned or mentioned by** [@alibabaqwen](/creator/undefined) [@furaidosu](/creator/undefined) [@merjicai](/creator/undefined) [@alitongyilab](/creator/undefined) [@meituanlongcat](/creator/undefined) [@stepfunai](/creator/undefined) [@dx8152](/creator/undefined) [@kimimoonshot](/creator/undefined) [@nhldesktop](/creator/undefined) [@zhihufrontier](/creator/undefined) [@tencenthunyuan](/creator/undefined) [@alibabacloud](/creator/undefined) [@pranaysuyash](/creator/undefined) [@ali_tongyilab](/creator/undefined) [@barrakali](/creator/undefined) [@1328](/creator/undefined) [@github](/creator/undefined) [@baiduinc](/creator/undefined) [@isaifai](/creator/undefined) [@crazytoolman](/creator/undefined) **Top assets mentioned** [Robot Consulting Co., Ltd. (LAWR)](/topic/robot) [Alphabet Inc Class A (GOOGL)](/topic/$googl) ### Top Social Posts Top posts by engagements in the last [--] hours "@Alibaba_Qwen @github yes we are still here😉🙌🙌" [X Link](https://x.com/MaaSAI42/status/1831592777714692396) 2024-09-05T07:18Z [--] followers, [---] engagements "🚀 We have launched nano-banana Studio You can try it on ModelScope for FREE🔗Link in the first comment 🔥 Here are [--] tips for beginners to make prompts clear stable and controllable👇 [--] Lock what must not change Tell the model what to preserve: character pose expression colors. Example: Keep the main character unchanged. [--] Start with verbs Be direct: remove / add / replace / turn into. Avoid vague phrasing like adjust a bit. [--] Be ultra-specific Describe object + color + light + material + style. Example: Change the background to pure white no shadows. [--] Layer your prompt Structure = Subject" [X Link](https://x.com/ModelScope2022/status/1960612308214013968) 2025-08-27T07:56Z [---] followers, [---] engagements "🚀 Exciting update from @Kimi_Moonshot Kimi K2-0905 brings enhanced coding 256k context and better agent integration. Check out the impressive benchmarks vs Claude Sonnet [--] and try the turbo API for 100% tool-call accuracy. 🔗 https://modelscope.cn/models/moonshotai/Kimi-K2-Instruct-0905 Kimi K2-0905 update 🚀 - Enhanced coding capabilities esp. front-end & tool-calling - Context length extended to 256k tokens - Improved integration with various agent scaffolds (e.g. Claude Code Roo Code etc) 🔗 Weights & code: https://t.co/83sQekosr9 💬 Chat with new Kimi https://t.co/mkOuBMwzpw" [X Link](https://x.com/ModelScope2022/status/1963809724735934555) 2025-09-05T03:41Z [---] followers, [---] engagements "🎨✨ Say hello to HunyuanImage-2.1 your new open-source AI art powerhouse: - 2K resolution at lightning speed - Multi-language prompts (English/Chinese) - Flexible aspect ratios (1:1 16:9 9:16 4:3 3:4 3:2 2:3) - Smart PromptEnhancer for effortless detail & style - Quality on par with Seedream3.0 and Qwen-Image 🔗 Check it out: #hunyuanimage #AIart https://modelscope.cn/models/Tencent-Hunyuan/HunyuanImage-2.1/summary https://modelscope.cn/models/Tencent-Hunyuan/HunyuanImage-2.1/summary" [X Link](https://x.com/ModelScope2022/status/1965341735317573918) 2025-09-09T09:09Z [---] followers, [---] engagements "🚀Huge congrats to @Alibaba_Qwen Qwen3-Next is seriously innovative. As community members put it: Qwen just set the new baseline. Might be the first shot in the race for hyper-complex hybrid-arch models. Lets break down what it brings: ⚡ Smaller model size same capability 💰 [--] cheaper GPU hours 📜 [--] longer context 🥊 Beats Gemini-2.5-Flash-Thinking Full thread" [X Link](https://x.com/ModelScope2022/status/1966422045048348769) 2025-09-12T08:42Z [---] followers, [---] engagements "1 Hybrid Architecture - 75% of layers use efficient linear attention (Gated DeltaNet) - 25% keep enhanced standard attention balances long-context efficiency & recall fidelity Enhancements (standard-attention layers only): - Output gating to curb low-rank bottlenecks - Head dimension upsized: [---] [---] - RoPE applied only to first 25% of positions per head boosts length extrapolation" [X Link](https://x.com/ModelScope2022/status/1966423020312166883) 2025-09-12T08:45Z [---] followers, [--] engagements "2 Extreme Sparse MoE - 80B total parameters but only 3B activated per inference - Architecture: [---] experts + [--] routing experts + [--] shared expert Global load balancing decouples parameter scale from inference cost max efficiency without losing performance" [X Link](https://x.com/ModelScope2022/status/1966423375712244013) 2025-09-12T08:47Z [---] followers, [--] engagements "4 Native Multi-Token Prediction (MTP) - Provides an MTP module with high Speculative Decoding acceptance rates - Enhances backbones overall performance - Optimizes multi-step inference via inference-consistent multi-step training further improves Speculative Decoding in practical scenarios" [X Link](https://x.com/ModelScope2022/status/1966423793397874792) 2025-09-12T08:49Z [---] followers, [--] engagements "@Baidu_Inc has launched ERNIE X1.1 claiming it outperforms DeepSeek-R1-0528 and matches GPT-5 and Gemini [---] Pro in performance. Early users report noticeable improvements especially in reasoning. Worth a trycheck it out and test it in different scenarios ERNIE X1.1 just launched Unveiled at Wave Summit [----] our latest reasoning model reduces hallucinations improves instruction following and delivers strong agentic capabilities. ✅ Achieves significant performance gains over ERNIE X1 with factual accuracy up 34.8% https://t.co/6LkjZqvRS0 ERNIE X1.1 just launched Unveiled at Wave Summit 2025" [X Link](https://x.com/ModelScope2022/status/1967560657286631656) 2025-09-15T12:06Z [---] followers, [--] engagements "1/5 The Wan-Animate paper is out With just a single character image and a reference video Wan-Animate can seamlessly replace the original characterperfectly capturing the targets expressions movements and the scenes lighting and color tones. The result is AI-generated video that feels natural consistent and free of visual artifacts. So how does Wan-Animate achieve this👇👇👇 paper: #Wan #WanAnimate https://arxiv.org/abs/2509.14055v1 https://arxiv.org/abs/2509.14055v1" [X Link](https://x.com/ModelScope2022/status/1968641432635216204) 2025-09-18T11:41Z [---] followers, [---] engagements "5/5 Finally the Wan team showcased a diverse range of results produced by Wan-Animate highlighting its stability controllability and state-of-the-art performance" [X Link](https://x.com/ModelScope2022/status/1968642251724034255) 2025-09-18T11:44Z [---] followers, [---] engagements "🚀ModelScope FlowBench is now live A local-cloud collaborative platform for text / image / 3D / video / audio workflows. ✅Free cloud compute run QwenImage Wan2.2 Flux SD and more with zero GPU hassle. ✅ One-click workflow cloning copy pro pipelines instantly. Change outfits swap background figurine effects all in a single prompt. ✅ 10000+ built-in LoRAs choose switch and apply with ease. ✅ Drag-and-drop nodes + real-time preview build complex pipelines with generation editing and pose control all in one. 💻 For Free jump in today Local client download in the comments #Aiart #AITools" [X Link](https://x.com/ModelScope2022/status/1971130987368386819) 2025-09-25T08:33Z [---] followers, [----] engagements "@nhl_desktop thanks for letting me know try links below😎 🍎 💻 https://cdn-muse-cn-1.modelscope.cn/flowbench/release/win/FlowBench-0.5.0-beta-20250923-2111.exe https://cdn-muse-cn-1.modelscope.cn/flowbench/release/mac/FlowBench-0.5.0-beta-20250923-2111.dmg https://cdn-muse-cn-1.modelscope.cn/flowbench/release/win/FlowBench-0.5.0-beta-20250923-2111.exe https://cdn-muse-cn-1.modelscope.cn/flowbench/release/mac/FlowBench-0.5.0-beta-20250923-2111.dmg" [X Link](https://x.com/ModelScope2022/status/1971137548656443639) 2025-09-25T08:59Z [---] followers, [--] engagements "@nhl_desktop not yet but we'll certainly consider it" [X Link](https://x.com/ModelScope2022/status/1971139339280318603) 2025-09-25T09:06Z [---] followers, [--] engagements "🎉ModelScope DiffSynth-Studio Hits [-----] GitHub Stars Huge thanks for your incredible support 🔥 We've completely redesigned the inference and training pipelines for leading Diffusion Models like Qwen-Image Wan Hunyuan and Flux. This delivers efficient dynamic VRAM management and highly flexible model training making it easy for consumer-grade GPUs to handle large models 🌟 Key Highlights: - The ModelScope Civision has leveraged DiffSynth-Studio to generate 3000+ Qwen-Image LoRA models. - Our versatile training framework has rapidly incubated multiple Qwen-Image extension models significantly" [X Link](https://x.com/ModelScope2022/status/1972237495833395475) 2025-09-28T09:50Z [---] followers, [---] engagements "DeepSeek-V3.2-Exp is hereopen-source and blazing fast 🚀 - Tech: Powered by DeepSeek Sparse Attention delivering huge speedups on long-context training & inference with no quality tradeoff. - Performance: Matches V3.1-Terminus across top public benchmarks. - Price: Serving costs slashedAPI now over 50% cheaper effective today 🔗Get the model and try it out on ModelScope: 🔗Paper link: #DeepSeek #LLM https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2-Exp" [X Link](https://x.com/ModelScope2022/status/1972618946189893762) 2025-09-29T11:06Z [---] followers, [---] engagements "1/4 🤯 The Embodied AI "ChatGPT Moment" is 1-2 years away. But what's the path The ModelScope Embodied AI Developer Forum saw experts from DAMO Academy BAAI OpenLoong Ant Group Manycore X Square Robot Linkerbot and more unveil their roadmap for the future. The consensus: Data is the new bottleneck and open source is the solution. Top [--] breakthroughs and key takeaways in this thread 👇🧵" [X Link](https://x.com/ModelScope2022/status/1976947653796937900) 2025-10-11T09:47Z [---] followers, [---] engagements "4/4 The Community & Commercial Tipping Point: 🚀 OpenLoong: Launched China's first full-stack open-source full-size Humanoid Robot (1.85m [--] DoF) & the low-cost NanoLoong educational kit. 💸 Consumer Price Point: Experts predict a reliable multi-task robot at the [----] RMB price point could be the catalyst for mass adoption" [X Link](https://x.com/ModelScope2022/status/1976961823552479317) 2025-10-11T10:43Z [---] followers, [---] engagements "@iSaif_ai @Merjic_AI @Alibaba_Qwen Yes but make sure to choose Qwen-Image-Edit v1" [X Link](https://x.com/ModelScope2022/status/1979171508162355392) 2025-10-17T13:03Z [---] followers, [---] engagements "🔥 Shanghai AI Laboratory open-sources InternVL3.5 🚀 ✨ Highlights: [--] model sizes (1B 241B params) dense + MoE first open multimodal LLM built on GPT-OSS. 241B model tops benchmarks: MMStar [----] OCRBench [----] beats GPT-5. AIME25 [----] MMLU-Pro [----] SOTA multimodal reasoning. Cascade RL boosts avg reasoning +16 pts; 241B model hits [----] beating Claude-3.7-Sonnet. New ViR & DvD slash latency [--] (369 ms [--] ms); Flash model keeps 100% perf at 50% seq len. Upgraded agents: GUI embodied AI SVG. New SOTA on ScreenSpot (92.9) VSI-Bench (69.5) SGP-Bench (70.6). 💻 Download here: 🌐 Try it: #Multimodal" [X Link](https://x.com/ModelScope2022/status/1962469272770519306) 2025-09-01T10:55Z [----] followers, [---] engagements "🤯 [---] Token/S on a MacBook Yes you read that right Shaohong Chen just fine-tuned the Qwen3-0.6B LLM in under [--] minutes using Apple's MLX framework. This is how you turn your MacBook into a serious LLM development rig. A step-by-step guide and performance metrics inside 🧵 https://mp.weixin.qq.com/s/5sIc41fE7LkEv4ytJ7j8Eg https://mp.weixin.qq.com/s/5sIc41fE7LkEv4ytJ7j8Eg" [X Link](https://x.com/ModelScope2022/status/1977706364563865805) 2025-10-13T12:01Z [---] followers, 13.6K engagements "🚀Qwen3-VL Ultimate Guide: FREE API & Fine-Tuning Tutorial We've packed two must-have skills into one easy-to-follow guide: [--] Free Qwen3-VL API Access: Start building without spending a dime. [--] ms-swift Fine-Tuning Tutorial: Customize Qwen3-VL easily with your own data using MS-Swift. Save the long image (it's a 💰 goldmine) #qwen3VL #qwen #msswift" [X Link](https://x.com/ModelScope2022/status/1978333110640296068) 2025-10-15T05:32Z [---] followers, [----] engagements "1/3 🚀 Introducing Face-to-Photo by DiffSynth-Studio & @Merjic_AI Transform ordinary face photos into stunning high-fidelity portraits. It's now open-source 👍 Built on Qwen-Image-Edit @Alibaba_Qwen the Face-to-Photo model excels at precise facial detail restoration. Unlike previous models (e.g. InfiniteYou) it captures fine-grained facial features across angles sizes and positions producing natural aesthetically pleasing portraits. 🌟 Local deployment/usage tutorial in the thread 🧵 Model Try it: https://modelscope.cn/aigc/imageGenerationtab=advanced&imageId=17008179" [X Link](https://x.com/ModelScope2022/status/1979069636953149575) 2025-10-17T06:19Z [---] followers, 21.7K engagements "@Merjic_AI @Alibaba_Qwen 3/3 💻 Experience it in FlowBench: Use the Image-to-Image Generation node select Qwen-Image-Edit v1 + DiffSynth-Studio/Qwen-Image-Edit-F2P LoRA generate high-quality portraits with one click. https://cdn-muse-cn-1.modelscope.cn/flowbench/release/win/FlowBench-0.5.0-beta-20250929-1751.exe https://cdn-muse-cn-1.modelscope.cn/flowbench/release/mac/FlowBench-0.5.0-beta-20250929-1751.dmg https://cdn-muse-cn-1.modelscope.cn/flowbench/release/win/FlowBench-0.5.0-beta-20250929-1751.exe" [X Link](https://x.com/ModelScope2022/status/1979070317533499433) 2025-10-17T06:21Z [---] followers, [---] engagements "@crazytoolman @Merjic_AI @Alibaba_Qwen we'd love to get your feedback on this😀" [X Link](https://x.com/ModelScope2022/status/1979093176137703729) 2025-10-17T07:52Z [---] followers, [---] engagements "1/10 🤖 Deep Research Agents are redefining search Driven by OpenAI & Google the Autonomous Research Paradigm is officially mainstream. But what does a top-tier agent look like under the hood This deep dive answers: - Core definitions & capabilities of Deep Research - Frontier architecture & iterations - Engineering insights & commonalities of mainstream architecture and design We break down Architectures Core Techniques (SFT/RL) and lessons from Perplexity DeerFlow and Tongyi DeepResearch 👇🧵 Full Report (Written by LiGongsheng from Modelscope Team): #DeepResearch #AIAgent" [X Link](https://x.com/ModelScope2022/status/1980539483989946818) 2025-10-21T07:39Z [---] followers, [---] engagements "3/10 Deep Research Agent architectures fall into two categories based on LLM autonomy: Static Workflows (fixed human-defined pipelines) offer stability and are easier to implement but lack generalization; Dynamic Workflows (model-driven planning execution and reflection) provide superior flexibility and generalization for complex tasks but demand higher LLM capability and face instability challenges. A hybrid approach is often used in practice to balance stability and agility" [X Link](https://x.com/ModelScope2022/status/1980539847522845177) 2025-10-21T07:41Z [---] followers, [--] engagements "4/10 Dynamic workflows are split between Single-Agent (relying on a single LLM's long-context reasoning for planning/execution/reflection ideal for end-to-end RL e.g. Agent-R1) and Multi-Agent (allocating sub-tasks to specialized agents mimicking teamwork offering scalability e.g. deerflow). Multi-Agent systems face challenges in coordination and context management while Single-Agent designs require top-tier base models" [X Link](https://x.com/ModelScope2022/status/1980539936198836590) 2025-10-21T07:41Z [---] followers, [--] engagements "7/10 Four main techniques are employed to optimize DR Agents: Prompt Engineering (e.g. ReAct CoT) Supervised Fine-Tuning (e.g. Open-RAG AUTO-RAG) Reinforcement Learning (RL) (e.g. Agent-R1 WebThinker) and Non-parametric Continuous Learning (e.g. Case-Based Reasoning/CBR used in AgentRxiv). Specifically Tongyi DeepResearch uses a two-stage SFT+RL curriculum training to achieve state-of-the-art results on multiple benchmarks (e.g. HLE=32.9 BrowseComp-EN=43.4)" [X Link](https://x.com/ModelScope2022/status/1980540109381595238) 2025-10-21T07:42Z [---] followers, [--] engagements "8/10 Open-Source Deep Dive. Two open-source projects offering key lessons: Deep Research (17k stars): Uses a recursive search tree (custom depth/breadth) + LLM Distillation to beat context window limits. Great for cost control. DeerFlow (15.2k stars): A robust Multi-Agent system with Human-in-the-Loop (plan modification) + Global State Management for seamless context flow" [X Link](https://x.com/ModelScope2022/status/1980540232731889862) 2025-10-21T07:42Z [---] followers, [--] engagements "9/10 Key engineering insights from current DR Agent development: - Dynamic Adaptation: Continuously re-evaluate which parts of the workflow the LLM should autonomously manage as model capabilities improve; - Iterative Search: Design progressive search pipelines where new queries are generated based on previous "learnings/findings" to avoid redundancy; - Clean Context: Deliver structured "learnings/findings" instead of raw text to subsequent agents for better stability and lower cost; - Human-in-the-loop: Implement simple yet critical mechanisms for intent clarification and plan modification" [X Link](https://x.com/ModelScope2022/status/1980540324805325215) 2025-10-21T07:43Z [---] followers, [--] engagements "🤯 MIT License + 5-Minute Coherence + 10x Speed Boost. Meituan just open-sourced LongCat-Video (13.6B) a SOTA video base model that's a serious contender for the World Model race. 🎥 The Breakthrough: Natively generates up to 5-minute continuous videos by pre-training on continuation tasks solving quality decay and physics adherence. Its better than most open-source models and even challenges Google's Veo3 in core metrics. ⚡ The Efficiency: Get 10x faster inference on a single H800. This is powered by Block-Sparse Attention (BSA) which cuts computation to 10% and sampling distillation (50" [X Link](https://x.com/ModelScope2022/status/1982760980200235250) 2025-10-27T10:47Z [----] followers, [---] engagements "1/3 🚀 Open-Source #1 + 10B Activation + Agent Mastery. Introducing MiniMax-M2 the MoE model (230B total 10B activated) built specifically to master complex Coding & Agent workflows. 🏆 Ranked #1 among global open-source models by Artificial Analysis for general intelligence. 👇 Why this lightweight powerhouse is the most efficient choice for your next Agent project. 🧵" [X Link](https://x.com/ModelScope2022/status/1982772055637713089) 2025-10-27T11:31Z [---] followers, [---] engagements "3/3 End-to-End Coding & True Agent Intelligence. M2 is not just a coder; it's a developer agent. It handles multi-file projects auto-fixes code via test verification and excels on SWE-Bench and Terminal-Bench. For Agent work M2 masters complex toolchains (Shell Python Browser) and showed the crucial ability for self-correction and recovery in the BrowseComp benchmark" [X Link](https://x.com/ModelScope2022/status/1982772295820255338) 2025-10-27T11:32Z [---] followers, [---] engagements "New SOTA GUI Grounding model UI-Ins is here from @Ali_TongyiLab & RUC🚀It's the first to treat user commands as dynamic reasoning pathways enabling human-like multi-perspective intent understanding. 📊 Achieves a 74.1% Task Success Rate in AndroidWorld surpassing Gemini [---] Computer Use (69.7%) 🌍 Open Source (7B/32B models) with full SFT/RL code. Dive in👇🧵 🤖 UI-Ins-7B: UI-Ins-32B: 📄 arXiv: https://modelscope.cn/papers/2510.20286 https://modelscope.cn/models/Tongyi-MiA/UI-Ins-32B https://modelscope.cn/models/Tongyi-MiA/UI-Ins-7B https://modelscope.cn/papers/2510.20286" [X Link](https://x.com/ModelScope2022/status/1983892229082247454) 2025-10-30T13:42Z [---] followers, 25.6K engagements "4/4 Size Isn't Everything: The 7B Emergent Win On MMBench-GUI L2s Advanced subset (tasks requiring implicit intent) the UI-Ins-7B model beats Qwen2.5-VL-7B by 159% Why the huge jump Structured reasoning generic free-form "lets think step by step" CoT which actually hurt grounding performance. The UI-Ins 4-view reasoning framework proves how you reason matters more than how large your parameter count is. 🧠" [X Link](https://x.com/ModelScope2022/status/1983892641143255103) 2025-10-30T13:43Z [---] followers, [---] engagements "🚀 Introducing LongCat-Flash-Omni a 560B-parameter (27B activated) open-source omni-modal MoE model excelling at real-time audio-visual interaction. Built on LongCat-Flashs high-performance shortcut-connected MoE architecture with zero-computation experts plus efficient multimodal perception & speech reconstruction modules. ✅ Unified offline multimodal understanding + real-time AV interaction ✅ 128K-token context window ✅ Trained via curriculum-inspired progressive strategy strong in both uni- and multi-modal tasks ✅ MIT-licensed full weights open-sourced 📊 SOTA results across OmniBench" [X Link](https://x.com/ModelScope2022/status/1984547272617837018) 2025-11-01T09:05Z [----] followers, 12.6K engagements "🚨@Kimi_Moonshot just open-sourced Kimi Linearthe FIRST hybrid attention architecture that BEATS full attention under fair comparisons. No training tricks. No parameter inflation. Just better engineering. Why this changes everything for LLM inference 👇 ✨ CORE INNOVATION: Kimi Delta Attention (KDA) Channel-wise gating (not head-wise) for precise memory management Specialized DPLR transition matrices for hardware efficiency First production-ready linear attention that doesn't sacrifice quality 100% PyTorch no custom CUDA required ⚡ REAL-WORLD GAINS (48B param model): 128K context: [----] faster" [X Link](https://x.com/ModelScope2022/status/1985223059801096552) 2025-11-03T05:50Z [---] followers, [----] engagements "🎨 Big news from ModelScope 🎨 One of our community developers just trained the Qwen-Image-Edit-2509 LoRA using ModelScopes no-code training tools and its now trending on Hugging Face 🚀 Were thrilled to see creatives turning their ideas into powerful LoRA models with Qwen-Image-Edit. Even better is now open globally 🌍Free training for Qwen-Image and other T2I models is live with Qwen-Image-Edit LoRA training support coming later this month. 💫Try it now: https://modelscope.ai/civision/modelTraining http://ModelScope.ai https://modelscope.ai/civision/modelTraining http://ModelScope.ai" [X Link](https://x.com/ModelScope2022/status/1985595997859561537) 2025-11-04T06:32Z [---] followers, [----] engagements "Thanks @CGTNOfficial for showcasing the incredible open-source AI revolution From mind-reading glasses to knowledge-hungry robots the future is already here. ModelScope is one of the open communities where this revolution begins with every developer who dives in. Watch the video and join the movement When Code Has No Secrets Who Defines the Future #PulseofModernization #APEC2025 https://t.co/te0QPqS0W6 When Code Has No Secrets Who Defines the Future #PulseofModernization #APEC2025 https://t.co/te0QPqS0W6" [X Link](https://x.com/ModelScope2022/status/1985986499112743103) 2025-11-05T08:24Z [---] followers, [---] engagements "1/4Just fine-tuned Qwen3-8B on Huawei Ascend NPU for quant finance coding tasks 🔥 🤖Inspired by RD-Agentinstead of calling cloud LLMs to convert quant formulas code Yuan Guo fine-tuned Qwen3-8B to do it locally: faster cheaper and purpose-built for finance. Built a specialized model that converts complex LaTeX quant factor formulas directly into production-ready Python functions. Achieved +20 pts over base model on code correctness robustness & style.🚀 Key stack: Huawei Ascend 910B3 (64GB) LLaMA Factory @llamafactory_ai + LoRA SwanLab for experiment tracking MindNLP for deployment Best" [X Link](https://x.com/ModelScope2022/status/1986293821202960704) 2025-11-06T04:45Z [---] followers, [---] engagements "Special Release Who Codes the Future New data reveals the staggering scale of Chinas AI ecosystem: 9.4M+ developers 30M+ open-source projects 346K+ AI companies in H1 [----]. 🚀This isnt just growthits a full-stack innovation machine shaping global trends. D-3 until The Next Wave: Who Codes the Future The AI Developer Ecosystem Report by @ModelScope2022 x @ZhihuFrontier Every line of code carries a worldview. Every developer a choice about the future. #AI #Developers #ModelScope #Zhihu #ChinaAI" [X Link](https://x.com/ModelScope2022/status/1986624576454074507) 2025-11-07T02:39Z [----] followers, [---] engagements "🔥 BREAKING: @StepFun_ai just open-sourced Step-Audio-EditXthe world's first LLM-grade audio editing model 🎧✨ Control emotion speaking style AND paralinguistic elements (breaths laughs sighs) through simple text prompts. Zero-shot TTS with multilingual support (EN/CN + dialects) in a unified 3B-parameter architecture. ✅ Single GPU deployable ✅ Apache [---] license ✅ Iterative editing capabilities No more patchwork pipelinesaudio creation just got an LLM upgrade. Model: GitHub: #AI #AudioAI #OpenSource #LLM #GenerativeAI #SpeechTech #StepFunAI http://github.com/stepfun-ai/Step-Audio-EditX" [X Link](https://x.com/ModelScope2022/status/1987055393001066687) 2025-11-08T07:11Z [---] followers, [----] engagements "🚀 Meet @Meituan_LongCat 's UNO-Benchthe first unified benchmark that finally cracks the black box of omni-modal AI. The breakthrough: UNO-Bench reveals a compositional law: 🔹 Weak models hit a bottleneck 🔹 Strong models get superlinear synergy (P_omni (P_audio P_visual)2.19) ✅ [----] human-curated omni samples (98% cross-modal solvable) ✅ Novel multi-step open-ended QA + 95% accurate auto-grading ✅ 90% faster eval with 98% consistency vs. [--] public benchmarks If youre building or evaluating multimodal modelsthis is your new gold standard. 🔗 #AI #Multimodal #Benchmark #LLM #OmniModel" [X Link](https://x.com/ModelScope2022/status/1987427794264203604) 2025-11-09T07:51Z [---] followers, [----] engagements "🚀 ModelScope just launched Mcore-Bridge making high-performance Megatron training as easy as Transformers. Train MoE models up to 10x faster than DeepSpeed with seamless integration into vLLM SGLang & Hugging Face ecosystems. No more complex setup no manual weight conversion just plug & play with native safetensors support. 🔥 Key highlights: ✅ 10x speedup on MoE models (e.g. Qwen3-MoE) ✅ Native safetensors I/O no format hassles ✅ Full LoRA support + bidirectional PEFT compatibility ✅ Smooth Megatron vLLM weight sync for RLHF (GRPO/GKD) ✅ Python API & CLI for flexible debuggable workflows ✅" [X Link](https://x.com/ModelScope2022/status/1987818514217136468) 2025-11-10T09:44Z [---] followers, [----] engagements "🔥 3B active params. Full multimodal thinking. ERNIE-4.5-VL-28B-A3B-Thinking is here a lightweight vision-language model with deep reasoning image zoom search grounding & video understanding. ✅ Apache [---] Commercial use allowed ✅ ModelScope vLLM FastDeploy ready ✅ SFT/LoRA support via ERNIEKit Outperforms expectations across visual reasoning STEM and real-world tasks all at minimal cost. 👉 Try it now: #AI #MultimodalAI #LLM #OpenSource #AIAgent #ComputerVision #PaddlePaddle https://modelscope.cn/models/PaddlePaddle/ERNIE-4.5-VL-28B-A3B-Thinking" [X Link](https://x.com/ModelScope2022/status/1988196074381537446) 2025-11-11T10:44Z [---] followers, [----] engagements "🧠 Turn expertise into plug-and-play AI skills. 🚀MS-Agent now implements Anthropics Agent Skills protocol a modular way to give agents domain knowledge via structured folders of instructions scripts and resources. ❗Why it matters As agents grow more capable we need composable portable and scalable ways to inject specialized knowledge beyond simple tool calling. Skills solve this by packaging human workflows as reusable onboarding guides for AI. 🔧 How it works - Skills live in folders with (YAML + Markdown) optional scripts/ resources/ - MS-Agent loads them on-demand using 4-level context" [X Link](https://x.com/ModelScope2022/status/1988230025540497678) 2025-11-11T12:59Z [---] followers, [---] engagements "Large models have made development easier and sped up progress but theyve also birthed a new role: the LLM Toilet Paper Engineer.🤣 Were now living in the 80-Point Crisis where everything almost works but no one dares to ship it. Executives see the magic. They dont see the mess behind it. Jinhui a contributor from @ModelScope2022 and @ZhihuFrontier shared his insights during a rencent panel -- and what do you think" [X Link](https://x.com/ModelScope2022/status/1988481773874147542) 2025-11-12T05:39Z [---] followers, [----] engagements "KookYan just released Kook_Qwen_zshx_v2 a powerful LoRA built on Qwen-Image🚀 👇Example prompts are in the comments below Its designed to SMASH the boundary between real-life photorealism and CG fantasy. 🧙This powerful LoRA masters the "Real-life cinematic quality + pure fantasy scene" fusion serving as a massive shortcut for film concept artists game developers and virtual content creators looking to produce "cinematic real-fantasy fusion works" efficiently 💡 Key Features for Creators: - Realism in Fantasy: Achieve the "real person texture + fantasy scene" fusion without the typical" [X Link](https://x.com/ModelScope2022/status/1988947604944928851) 2025-11-13T12:30Z [---] followers, [----] engagements "8. Asian female real person. "Mandatory specified art style": "Realistic style photo taken with a SLR camera" "Lens": "Medium shot equivalent 85mm focal length shallow depth of field eye-level perspective slightly front-left angle highlighting the subject's silhouette and details" "Composition": "Subject slightly right of center vertical composition foreground is the upper body of the character background is a rough rock wall; character occupies about two-thirds of the image height cropped below the shoulders emphasizing the torso and costume details; background blurred but retaining texture" [X Link](https://x.com/ModelScope2022/status/1988948667437576574) 2025-11-13T12:34Z [---] followers, [---] engagements "9. A sexy girl dancing movements are large and beautiful wearing a yellow-green and gold colored Hanfu off-the-shoulder exquisite hair accessories full body beautiful leg lines low-cut sexy Da Vinci color grading transparency real person realistic background is a dark partially light-transmitting golden palace slight low angle slightly tilted lens delicate fair skin with a dewy luster cold and arrogant eyes natural messy strands of hair falling down subtle reflection on hair strands black smooth long hair proud figure full and upright breasts golden characters "Chang Le Wei Yang" (Eternal Joy" [X Link](https://x.com/ModelScope2022/status/1988948744633741441) 2025-11-13T12:35Z [---] followers, [---] engagements "11. surreal realistic_photo photography high_contrast hyper_detailed low_saturation "caption": "A red-haired woman wearing Hanfu a large-sleeved Hanfu of the Wei and Jin style holding a long scepter the scepter is taller than her the head of the scepter is a copper ring strings of small bells hang from the copper ring swaying in the wind as if hearing the crisp sound of the bells. She faces the camera tilting her head smiling slightly. The skirt is long and extends to the ground. Emphasizing the messy beauty of dynamic hair strands extremely high visual appeal Rembrandt light reflection" [X Link](https://x.com/ModelScope2022/status/1988949094191239550) 2025-11-13T12:36Z [---] followers, [---] engagements "🔥 Big shoutout to @Eigen_AI_Labs for this release Our DiffSynth-Studio team saw it and theyre very happy. 😄 Meet Eigen-Banana-Qwen-Image-Edit: ⚡ 4-step pro-quality edits (down from 50+) 🖋 Pixel-perfect text + semantic control 🧩 Open LoRA weights runs natively in DiffSynth-Studio & Diffusers 🛠 Built on EigenPlatform + Pico-Banana-400K eigen-banana-qwen-image-edit Eigen-Banana-Qwen-Image-Edit is a LoRA (Low-Rank Adaptation) checkpoint for the Qwen-Image-Edit model optimized for fast high-quality image editing with text prompts. This model enables efficient text-guided image transformations" [X Link](https://x.com/ModelScope2022/status/1989264113256792537) 2025-11-14T09:28Z [---] followers, [----] engagements "🚀Meet MemOS: the first Memory Operating System for LLMs. Open source. Built for agents with long-term memory. ✅ Persistent searchable evolving memories ✅ Standard MCP I/O no SDKs no plugins ✅ Works with ModelScope 14.9k+ calls in [--] week 🔗Try it live: 🔗Star on GitHub: Let memory become AIs new compute layer. #AI #LLM #MemOS #MCP http://github.com/MemTensor/MemOS https://modelscope.cn/mcp/servers/MemTensor/MemoryOperatingSystem http://github.com/MemTensor/MemOS https://modelscope.cn/mcp/servers/MemTensor/MemoryOperatingSystem" [X Link](https://x.com/ModelScope2022/status/1990325195576390014) 2025-11-17T07:44Z [---] followers, [----] engagements "1/4 🚀 Meet AgentEvolver from @Ali_TongyiLab : AI agents that learn like humans no hand-labeled data needed. A 7B model beats a 14B baseline. 57.6% success rate vs. 29.8% with 50% fewer parameters. The team have open-sourced everything: 🔗 Code: 📄 Paper: No more costly RL loops or random exploration. ✅ Self-questioning generates its own tasks ✅ Self-navigating reuses experience smarter paths ✅ Self-attributing rewards smart actions not just luck Built on LLMs optimized for efficiency. Faster adaptation. Less data. More autonomy. "The future of AI agents isnt about bigger models. Its about" [X Link](https://x.com/ModelScope2022/status/1990707909517586560) 2025-11-18T09:05Z [---] followers, [----] engagements "3/4 1Self-Questioning Instead of waiting for humans to define tasks AgentEvolvers LLM looks at its environment and asks: What can I try next It autonomously generates novel meaningful challenges no manual dataset needed. Think of it as curiosity encoded. 2Self-Navigating Every failure every success its not wasted. The agent summarizes past experiences into reusable mental maps then blends them with fresh exploration. No more repeating the same mistakes. It learns from history not just trial-and-error. 3Self-Attributing In long tasks you rarely know which step led to success. AgentEvolver runs" [X Link](https://x.com/ModelScope2022/status/1990708370891026800) 2025-11-18T09:07Z [---] followers, [---] engagements "🔥The LoRAs built on Qwen-Image-Edit-2509 are blowing upand have dominated Hugging Faces Trending list. All were trained on ModelScope and are available via free API: 📷 Multiple-angles rotate product shots like a [---] camera 💡 Light_restoration erase shadows enhance lighting naturally 🛒 Fusion & White_to_Scene seamlessly place products into real-world scenes (no Photoshop needed) 🕯 Relight simulate soft window glow studio lighting and more 👥 Multiple-characters generate entire scenes with consistent characters from a single image 🔧 Workflow expand retouch virtual try-on pose control all" [X Link](https://x.com/ModelScope2022/status/1991116417765351833) 2025-11-19T12:08Z [----] followers, [----] engagements "🚀 @TencentHunyuan Just dropped: HunyuanVideo [---] the lightweight SOTA video gen model that runs on a 14GB consumer GPU. 8.3B params. DiT architecture. Realistic 510s 480p/720p videos. Upscale to 1080p. ✅ Mid-prompt control (English/Chinese) ✅ Image-to-video with perfect consistency ✅ Cinematic camera moves text in video physics-aware motion ✅ SSTA attention = 3x faster inference than 20B+ models No A100 needed. Just deploy & play. ModelScope: GitHub: https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5 https://www.modelscope.cn/models/Tencent-Hunyuan/HunyuanVideo-1.5" [X Link](https://x.com/ModelScope2022/status/1991735494804402584) 2025-11-21T05:08Z [----] followers, [----] engagements "Meet Nex by NEX-AGI a non-thinking model built for agents that crushes it in coding tool use and roleplay 🚀 ✅ SOTA among open models on Tau2-Bench BFCL V4 GAIA2 ✅ Top-tier in frontend vibe coding and mini-program/backend dev (human eval confirmed) ✅ Plug-and-play with Claude Code Cursor etc. 👉 Expolore Nex: 🔥 Free for now on SiliconFlow: https://www.siliconflow.com/models/deepseek-v3-1-nex-n1 https://modelscope.cn/organization/nex-agi https://www.siliconflow.com/models/deepseek-v3-1-nex-n1 https://modelscope.cn/organization/nex-agi" [X Link](https://x.com/ModelScope2022/status/1992836661592629614) 2025-11-24T06:04Z [----] followers, [----] engagements "🔥 @TencentHunyuan just open-sourced HunyuanOCR a 1B-parameter end-to-end multimodal OCR model that beats Gemini Qwen-VL and commercial APIs across 10+ benchmarks. ✅ SOTA on OmniDocBench (94.1) ✅ Full document parsing: LaTeX formulas HTML tables reading-order Markdown ✅ 100+ languages 14EN/ZH photo translation (ICDAR2025 small-model champ) ✅ JSON field extraction (receipts IDs) bilingual subtitle extraction ✅ Single inference no cascades lightweight & production-ready All with just 1B params 🚀 Try it now: 🔗 🔗 https://github.com/Tencent-Hunyuan/HunyuanOCR" [X Link](https://x.com/ModelScope2022/status/1993223757284970997) 2025-11-25T07:42Z [----] followers, [----] engagements "🚀Bytedance-Research just dropped: VeAgentBench a new benchmark dataset to evaluate real-world agentic capabilities ✅ [---] open Qs (484 total) across legal finance edu & personal assistant ✅ Tests tool use RAG memory & multi-step reasoning ✅ Comes with runnable agents built on veADK (Volcengines full-stack agent framework) Perfect for devs pushing agentic AI beyond toy demos. 🔗 Dataset: 📜 CC BY-NC [---] (non-commercial) https://modelscope.cn/datasets/bytedance-research/veAgentBench/summary https://modelscope.cn/datasets/bytedance-research/veAgentBench/summary" [X Link](https://x.com/ModelScope2022/status/1993308962058362946) 2025-11-25T13:21Z [----] followers, [----] engagements "Introducing Z-Image: a powerful highly efficient 6B-parameter image generation model with three specialized variants 🎨 🚀 Z-Image-Turbo Distilled for speed & quality Just [--] NFEs matches or beats top competitors ⚡Sub-second latency on H800 GPUs Runs smoothly on 16GB VRAM consumer cards Photorealism + bilingual text (EN/ZH) + strong instruction following Try it now: 🔗 Demo: 🔗 Model: https://modelscope.cn/models/Tongyi-MAI/Z-Image-Turbo https://modelscope.cn/aigc/imageGeneration https://modelscope.cn/models/Tongyi-MAI/Z-Image-Turbo https://modelscope.cn/aigc/imageGeneration" [X Link](https://x.com/ModelScope2022/status/1993861121565208985) 2025-11-27T01:55Z [----] followers, 21.5K engagements "Free Z-Image API is live 🎉 The open-source Z-Image Turbo model is now 100% free to call via API on ModelScope🔥 AND Free compute included ✨ How to use it: [--] Try it instantly: Go to Z-Image is the default model - Quick Mode: Fast results zero config - Advanced Mode: Full control for fine-tuned outputs [--] Integrate it anywhere: Visit 👉 copy the Python code and run it: - Locally or in Jupyter - In your own apps - With ComfyUI or other AI workflows Start generating now #ZImage_Turb https://modelscope.cn/models/Tongyi-MAI/Z-Image-Turbo https://modelscope.cn/aigc/imageGeneration" [X Link](https://x.com/ModelScope2022/status/1994300901914550400) 2025-11-28T07:02Z [----] followers, 50.2K engagements "🚀 Meet GELab-Zero-4B-preview a 4B multimodal GUI agent from @StepFun_ai fine-tuned for Android & runs locally on consumer hardware ✅ Zero-shot UI automation: taps swipes typing all from screenshots/video ✅ Handles complex cross-app tasks (food rides shopping social) ✅ No app-specific training needed works on unseen dynamic interfaces ✅ Plug-and-play infra: ADB deps recording/replay included Built on Qwen3-VL-4B-Instruct. Small model. Big autonomy. Local. Visual. General. 🔥 🤖Model: https://modelscope.cn/models/stepfun-ai/GELab-Zero-4B-preview" [X Link](https://x.com/ModelScope2022/status/1995326051363553501) 2025-12-01T02:56Z [----] followers, [----] engagements "DeepSeek-V3.2 is live 🚀 @deepseek_ai and its the most capable open-agent LLM yet. ✅ Matches GPT-5 on reasoning ✅ First open model to combine chain-of-thought + tool calling ✅ Trained on 85K+ synthetic agent tasks SOTA on SWE-Bench Tool Decathlon & more ✅ Built with DSA: [--] cheaper inference at 128K And the research variant V3.2-Speciale just won gold medals in IMO [----] IOI [----] ICPC WF on par with Gemini-3.0-Pro.🏆 📄 Tech Report: 🔗 Open weights on ModelScope: https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2 https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2-Speciale" [X Link](https://x.com/ModelScope2022/status/1995458642037145874) 2025-12-01T11:43Z [----] followers, 14K engagements "🔥 The Qwen-Image-Edit-2509 API is now live and 100% free to use ✨ Free inferences refresh daily 🎁 New users get free GPU hours on signup 🚀 Ready to build Try it on ModelScope: https://modelscope.cn/models/Qwen/Qwen-Image-Edit-2509 https://modelscope.cn/models/Qwen/Qwen-Image-Edit-2509 https://modelscope.cn/models/Qwen/Qwen-Image-Edit-2509 https://modelscope.cn/models/Qwen/Qwen-Image-Edit-2509" [X Link](https://x.com/ModelScope2022/status/1995767427448344871) 2025-12-02T08:10Z [----] followers, 31K engagements "🚀 Ovis-Image (7B) is live on ModelScope ✅Delivers frontier-level text renderingon par with 20B-class models like Qwen-Image and even competitive with GPT-4o on text-heavy tasks. ✅Sharp layout-aware output for posters banners logos UI mocks and infographics. ✅Runs fast and leandeployable on a single high-end GPU. Small model. Big text fidelity. 👉 https://modelscope.cn/models/AIDC-AI/Ovis-Image-7B https://modelscope.cn/models/AIDC-AI/Ovis-Image-7B" [X Link](https://x.com/ModelScope2022/status/1996040213337092215) 2025-12-03T02:14Z [----] followers, [----] engagements "🚀 Introducing Flowra by ModelScope & WULI The open-source engine behind FlowBench makes building AI workflows as easy as snapping LEGO bricks. ✅ Unified handling of images audio video & 3D ✅ DAG-based execution w/ smart caching & distributed scaling ✅ One-line ModelScope model integration ✅ Full dev toolkit: flowra create build debug deploy Turn your ML model into a visual node in minutes. ✨ No more dependency hell. No more blind debugging. 🔗 GitHub: 📥 FlowBench client: https://modelscope.cn/flowbench/download https://github.com/modelscope/flowra https://modelscope.cn/flowbench/download" [X Link](https://x.com/ModelScope2022/status/1996535420252020841) 2025-12-04T11:01Z [----] followers, [----] engagements "🔥 New I2V breakthrough: SteadyDancer generates ultra-coherent human animation videos while perfectly preserving the first frames identity. ✅ Solves identity drift & motion artifacts ✅ Uses smart conditional harmonization + pose-aware modules ✅ Trains faster + outperforms SOTA on FVD VBench SSIM No more uncanny valleyjust smooth controllable reference-accurate motion. 🤖 Model: 📄 Paper: https://modelscope.cn/papers/2511.19320 https://modelscope.cn/models/MCG-NJU/MCG-NJU-SteadyDancer-14B https://modelscope.cn/papers/2511.19320 https://modelscope.cn/models/MCG-NJU/MCG-NJU-SteadyDancer-14B" [X Link](https://x.com/ModelScope2022/status/1996775808166515182) 2025-12-05T02:57Z [----] followers, [----] engagements "🚀Introducing LongCat-Image from @Meituan_LongCat a groundbreaking open-source bilingual (Chinese-English) image generation foundation model 🐾 ✨ Just 6B params yet outperforms much larger models in efficiency & quality ✨ Best-in-class Chinese text rendering accurate stable and covers the vast majority of the Chinese lexile ✨ Striking photorealism powered by a novel data & training strategy Built for developers optimized for real-world use. Open accessible and multilingual from the ground up. 🤖Model: https://modelscope.cn/models/meituan-longcat/LongCat-Image" [X Link](https://x.com/ModelScope2022/status/1997170411508060490) 2025-12-06T05:05Z [----] followers, [----] engagements "🚀 NewBieAI-Lab drops NewBie-image-Exp0.1 a 3.5B open-source ACG-native DiT model built for precise fast and high-quality anime generation. ✅ 3.5B params (8GB VRAM friendly RTX [----] ✅) ✅ Dual text encoders: Gemma-3-4B-it + Jina CLIP v2 deep prompt understanding ✅ XML-structured prompts for per-character control (no more outfit swapping) ✅ FLUX.1-dev 16-ch VAE buttery skin fabric metal ✅ 20-step inference LoRA-friendly Apache-2.0 + non-commerical license ✅ Trained on 10M+ anime images w/ XML annotations rock-solid multi-character scenes 🤖Models: ⚡ 40% faster than 8B+ models yet handles" [X Link](https://x.com/ModelScope2022/status/1997543466587636209) 2025-12-07T05:47Z [----] followers, 43.2K engagements "🚀 Introducing Live Avatar: real-time streaming infinite-length interactive avatar video generation Powered by a 14B-parameter diffusion model [--] FPS on 5H800 GPUs (4-step sampling). Chat naturally via mic & camyour avatar responds instantly with lifelike video. Paired with Qwen3-Omni it enables fully autonomous face-to-face AI conversations. Watch two agents talkstreaming in real time for 10000+ seconds. 👀 Useful Resources👇 YouTube: Model: Blog: Paper: Code: https://github.com/Alibaba-Quark/LiveAvatar https://modelscope.cn/papers/2512.04677 https://liveavatar.github.io/" [X Link](https://x.com/ModelScope2022/status/1997872276256289012) 2025-12-08T03:34Z [----] followers, [----] engagements "Excellent work from @dx8152 🔗Qwen-Edit-2509-Light-igration: https://modelscope.ai/models/dx8152/Qwen-Edit-2509-Light-igration Love seeing what the community builds with @ModelScope2022 . @dx8152 just dropped a game-changing Light Migration LoRA for Qwen-Image-Edit-2509. It solves the "secondary lighting" headache perfectly. Incredible work. 👏 https://t.co/5IYTfR53uD https://modelscope.ai/models/dx8152/Qwen-Edit-2509-Light-igration Love seeing what the community builds with @ModelScope2022 . @dx8152 just dropped a game-changing Light Migration LoRA for Qwen-Image-Edit-2509. It solves the" [X Link](https://x.com/ModelScope2022/status/1997925749802864910) 2025-12-08T07:06Z [----] followers, [----] engagements "🚀This 80B LLM runs on 64GB RAMthanks to GGUF + smart quantization👏 Qwen3-Next-80B-A3B-InstructGGUFModelScope 64GBPCLLMhttps://t.co/me02jHdcrY Qwen3-Next-80B-A3B-InstructGGUFModelScope 64GBPCLLMhttps://t.co/me02jHdcrY" [X Link](https://x.com/ModelScope2022/status/1997939785101611400) 2025-12-08T08:02Z [----] followers, [----] engagements "🔥BREAKING DiffSynth-Studio just dropped: Qwen-Image-i2L the first open-source suite that turns a single image into a custom LoRA 🖼➡🧠 Choose your flavor: 🎨 Style pure aesthetic extraction (2.4B) 🧩 Coarse captures content + style (7.9B) ✨ Fine 1024x1024 detail booster (7.6B pairs with Coarse) ⚖ Bias aligns output with Qwen-Images native vibe (30M) Built on SigLIP2 + DINOv3 + Qwen-VL. Ready to train LoRAs from one image Its open. Its out. Go break something cool. 💥 🤖ModelScope: 💻Code:" [X Link](https://x.com/ModelScope2022/status/1998408862211441107) 2025-12-09T15:06Z [----] followers, 130.8K engagements "🚀 Good news for Z-Image Turbo users Weve open-sourced DistillPatch LoRA to restore true 8-step Turbo speed for any LoRA fine-tuned on Z-Image Turbo. 💡 The problem: Standard LoRA training breaks Z-Image Turbos accelerationimages turn blurry at steps=8 cfg=1 even though they look fine at steps=30. ✅ We tested [--] approaches: [--] Standard SFT high quality but slow [--] Diff-LoRA fast but inflexible [--] SFT + distillation great results heavy lift [--] SFT + DistillPatch (our recommendation) plug in one LoRA instantly get 8-step Turbo performance back 🚀 Perfect if you: Already have a Z-Image Turbo LoRA" [X Link](https://x.com/ModelScope2022/status/2000394797434163654) 2025-12-15T02:37Z [----] followers, 38.1K engagements "🚀 Tongyi FUN just leveled up This release brings major upgrades across TTS & ASR: 🧠 Fun-CosyVoice3 (0.5B) Open-sourced - Zero-shot voice cloning - Local deployment & secondary development ready Try Demo: Model(Chinese): Model(Multi-Language): ⚡ Fun-ASR-Nano (0.8B) Open-sourced - Lightweight Fun-ASR variant - Lower inference cost - Local deployment & custom fine-tuning supported Model: Open fast and production-ready. ✨ https://modelscope.cn/models/FunAudioLLM/Fun-CosyVoice3-0.5B-2512 https://modelscope.cn/models/FunAudioLLM/Fun-ASR-MLT-Nano-2512" [X Link](https://x.com/ModelScope2022/status/2000494474229879084) 2025-12-15T09:13Z [----] followers, 23.8K engagements "New Z-Image-Turbo control model released 🎉 More control blocks. Inpaint mode. Better details. - 6-block ControlNet - Trained from scratch on 1M images @1328 res - Canny / HED / Depth / Pose / MLSD - Best control_context_scale: 0.65-0.80 Model: Demo: Code: https://github.com/aigc-apps/VideoX-Fun https://modelscope.cn/studios/AI-ModelScope/ZIT-Controlnet https://modelscope.cn/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union https://modelscope.cn/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union https://github.com/aigc-apps/VideoX-Fun https://modelscope.cn/studios/AI-ModelScope/ZIT-Controlnet" [X Link](https://x.com/ModelScope2022/status/2000766764922401091) 2025-12-16T03:15Z [----] followers, 33.1K engagements "🚀Model & Demo are now live on our international site Model: - - - Demo: https://modelscope.ai/studios/FunAudioLLM/Fun-CosyVoice3-0.5B https://modelscope.ai/models/FunAudioLLM/Fun-CosyVoice3-0.5B-2512 https://modelscope.ai/models/FunAudioLLM/Fun-ASR-Nano-2512 https://modelscope.ai/models/FunAudioLLM/Fun-ASR-MLT-Nano-2512 https://modelscope.ai/studios/FunAudioLLM/Fun-CosyVoice3-0.5B https://modelscope.ai/models/FunAudioLLM/Fun-CosyVoice3-0.5B-2512 https://modelscope.ai/models/FunAudioLLM/Fun-ASR-Nano-2512 https://modelscope.ai/models/FunAudioLLM/Fun-ASR-MLT-Nano-2512" [X Link](https://x.com/ModelScope2022/status/2000864318758511044) 2025-12-16T09:43Z [----] followers, [---] engagements "🤖Meet Nanbeige4-3B from Boss Zhipina 3B-parameter LLM that outperforms Qwen3-32B on math (AIME) science (GPQA) and tool calling (BFCL-V4) while matching Qwen3-30B-A3B on human preference alignment (Arena-Hard-V2). How ✅ 23T tokens of ultra-curated data ✅ Fine-grained WSD scheduler ✅ 30M+ high-quality SFT instructions ✅ Multi-stage RL + innovative distillation (DPD) ✅ Chain-of-thought reconstruction & deliberative generation It even ranks top [--] on WritingBench & EQ-Bench3beating models 100x larger like GLM-4.5 and Deepseek-R1. All models + tech report now open-source: 🔗 Weights: 📄 Paper:" [X Link](https://x.com/ModelScope2022/status/2000890664486363362) 2025-12-16T11:28Z [----] followers, 25.8K engagements "🎉 MiMo-V2-Flash FREE API is now live on ModelScope The first major release since Fuli Luo joined Xiaomiand its built for real-world agentic AI. ⚡ MiMo-V2-Flash: an open high-performance MoE model with 309B total / 15B active parameters 256K context window 150+ tokens/s generation thanks to native Multi-Token Prediction 🔥 Key wins for developers: ✅ Hybrid attention (5:1 SWA + Global) [--] less KV cache full long-context recall ✅ 73.4% on SWE-Bench Verified new SOTA for open-source models ✅ Matches DeepSeek-V3.2 on reasoning but much faster in practice ✨ API-readyperfect for building smart" [X Link](https://x.com/ModelScope2022/status/2001479829095620906) 2025-12-18T02:29Z [----] followers, 16K engagements "🚀Qwen-Image-Layered is now live on ModelScopean incredible model that can intelligently decompose any image into multiple RGBA layers 🤖 Model: 🌟 Try Demo 📑 Paper: https://modelscope.cn/papers/2512.15603 https://modelscope.cn/studios/Qwen/Qwen-Image-Layered https://modelscope.ai/models/Qwen/Qwen-Image-Layered https://modelscope.cn/papers/2512.15603 https://modelscope.cn/studios/Qwen/Qwen-Image-Layered https://modelscope.ai/models/Qwen/Qwen-Image-Layered 🎨 Qwen-Image-Layered is LIVE native image decomposition fully open-sourced ✨ Why it stands out ✅ Photoshop-grade layering Physically" [X Link](https://x.com/ModelScope2022/status/2002038699878695146) 2025-12-19T15:29Z [----] followers, [----] engagements "Its still Sunday and Im already hyped for next week. 😊The mysterious Eastern power is dropping a trilogy of magic: 🔮 Next week: A secret image-generation model 🤖 Next week: A secret code-agent model 🎙 The week after: A secret voice model Buckle upthings are about to get interesting" [X Link](https://x.com/ModelScope2022/status/2002679068203028809) 2025-12-21T09:54Z [----] followers, 10.1K engagements "🚀 Meet GLM-4.7 your new coding partner supercharged Try it now on ModelScope free API access for a limited time ✅ +5.8% on SWE-bench (73.8%) ✅ +12.9% on Multilingual coding (66.7%) ✅ +16.5% on Terminal Bench [---] (41%) ✅ Smarter agent reasoning & tool use ✅ Better UI cleaner webpages sharper slides ✅ +12.4% on complex reasoning (HLE: 42.8%)From coding to creativity GLM-4.7 delivers across the board. 🤖 Model: https://modelscope.cn/models/ZhipuAI/GLM-4.7 https://modelscope.cn/models/ZhipuAI/GLM-4.7" [X Link](https://x.com/ModelScope2022/status/2003290346135187635) 2025-12-23T02:23Z [----] followers, [----] engagements "🚀 New on ModelScope: QwenLong-L1.5 is now fully open-source A 30B model (3B active params) that matches GPT-5 & Gemini-2.5-Pro in long-context reasoning. 🔥 Key wins: ✅ +31.7 pts on OpenAIs MRCR (128K context SOTA across all models) ✅ Matches Gemini-2.5-Pro on [--] major long-QA benchmarks ✅ +9.69 on CorpusQA +6.16 on LongBench-V2 How Three breakthroughs: [--] Synthetic data at scale: 14.1K long-reasoning samples from 9.2B tokens no human labeling. Avg. length: 34K tokens (max: 119K). [--] Stable RL training: Task-balanced sampling + Adaptive Entropy-Controlled Policy Optimization (AEPO) for reliable" [X Link](https://x.com/ModelScope2022/status/2003370363590226313) 2025-12-23T07:41Z [----] followers, 44.2K engagements "@Sean60133791259 @bdsqlsz Could I ask which country you are in Well test the speed. You can DM me" [X Link](https://x.com/ModelScope2022/status/2003648027375075799) 2025-12-24T02:04Z [----] followers, [--] engagements "Merry Christmas everyone 🎄✨ 🚀Introducing BEYOND REALITY Z IMAGE 1.0: a community-tuned checkpoint merging LoRA enhancements into Z-Image optimized for film-style portrait aesthetics and high-frequency texture detail think skin pores fabric weaves and wall surfaces rendered with striking realism. ✅ Keeps Zs photorealism intact ✅ Adds cinematic color depth & tonal nuance ✅ Significantly upgrades portrait + environmental textures ✅ Runs on 8GB VRAM (FP8) tested on laptops Note: Training focused on high-quality synthetic portrait photography so complex poses or non-human textures may vary." [X Link](https://x.com/ModelScope2022/status/2003737630245343345) 2025-12-24T08:00Z [----] followers, 13.5K engagements "We shouldve shouted this louder: Youre officially missing out if you aren't using Civision 🤫 🎨 FREE Image & Video Gen & FREE LoRA Training ✨ A UI so clean you dont need a PhD to master it. Weve got the heavy hitters ready: Z-Image Qwen-Image Qwen-Image-Edit Wan2.2 and more. All the power zero cost. Right here on ModelScope: 🚀✨ https://modelscope.ai/civision/imageGeneration https://modelscope.ai/civision/imageGeneration" [X Link](https://x.com/ModelScope2022/status/2004110641687417306) 2025-12-25T08:43Z [----] followers, [----] engagements "🚀New on ModelScope: ChenkinNoob-XL v0.2 is live Independently developed by the ChenkinNoob team as a fine-tuned extension of noobai-XL-1.1 v0.2 incorporates Danbooru data up to Nov [--] 2025enhancing performance while preserving the original aesthetic. ✨ Key improvements over v0.1: ✅ Stronger character consistency ✅ Refined details (hands accessories layered outfits) ✅ Higher prompt adherence 🎨 Faithful to the noob visual DNA: clean composition solid anatomy and harmonious color palettes. ⚙ Recommended config: CFG: [--] Steps: [----] Sampler: Euler a Resolution: [----] (e.g. [-------] 10241024)" [X Link](https://x.com/ModelScope2022/status/2004378356645814448) 2025-12-26T02:26Z [----] followers, 30.7K engagements "🚀 New on ModelScope: MiniMax M2.1 is open-source ✅ SOTA in 8+ languages (Rust Go Java C++ TS Kotlin Obj-C JS) ✅ Full-stack Web & mobile dev: Android/iOS 3D visuals vibe coding that actually ships ✅ Smarter faster 30% fewer tokens with lightning mode (M2.1-lightning) for high-TPS workflows ✅ Top-tier on SWE-bench VIBE and custom coding/review benchmarks ✅ Works flawlessly in Cursor Cline Droid BlackBox and more Its not just better code its AI-native development end to end. 🔗 Model: https://modelscope.cn/models/MiniMax/MiniMax-M2.1 https://modelscope.cn/models/MiniMax/MiniMax-M2.1" [X Link](https://x.com/ModelScope2022/status/2004462984698253701) 2025-12-26T08:03Z [----] followers, 17.3K engagements "@nnnnmailcom Hi there SDXL models should be supported. Could you let me know exactly what issue you're running into" [X Link](https://x.com/ModelScope2022/status/2005487313158279461) 2025-12-29T03:53Z [----] followers, [--] engagements "🚀400+ Free APIs are now available on We currently support: ✅ DeepSeek series ✅ Qwen3 series & Qwen-Image ✅ Z-Image-Turbo and many more Check it out: Go to - Models - Filter by API-Inference. (Currently available on PC only.) 👉 Give them a spin and let us know what you think ✉DM us for support. https://modelscope.ai/modelsfilter=inference_type&page=1&tabKey=task http://modelscope.ai http://ModelScope.ai http://modelscope.ai http://ModelScope.ai https://modelscope.ai/modelsfilter=inference_type&page=1&tabKey=task http://modelscope.ai http://ModelScope.ai http://modelscope.ai" [X Link](https://x.com/ModelScope2022/status/2005549668886040939) 2025-12-29T08:01Z [----] followers, 17.8K engagements "@blankbraindead oh I see. May I know which region you are in" [X Link](https://x.com/ModelScope2022/status/2005557772797215085) 2025-12-29T08:33Z [----] followers, [---] engagements "🚀 Tencent HunYuan open-sourced HY-MT1.5 A lightweight powerhouse for translation: ✨ 1.8B model: 0.18s latency runs offline on mobile & outperforms most commercial APIs. Even rivals Gemini-3.0-Pro at 90% quality ⚡ 7B model: Upgraded with WMT25-champion accuracy & fewer artifacts. 🌍 [--] languages + [--] Chinese dialects 🛠 Context-aware custom glossaries format-preserving output 💡 Perfect for edge + cloud deployments Models: https://modelscope.cn/collections/Tencent-Hunyuan/HY-MT15 https://modelscope.cn/collections/Tencent-Hunyuan/HY-MT15" [X Link](https://x.com/ModelScope2022/status/2005944042853667257) 2025-12-30T10:08Z [----] followers, 17K engagements "One of your New Year open-source gifts 🎨 is on its way" [X Link](https://x.com/ModelScope2022/status/2005968451538759734) 2025-12-30T11:45Z [----] followers, 10.7K engagements "🚀New on ModelScope: Qwen-Image-2512 is here 🎉 This version delivers a massive leap in realism and control: ✨ Photorealistic humans dramatically reduced AI look with accurate aging natural expressions and fine details (yes even individual hair strands) 🌿 Ultra-detailed nature fur foliage water and mist rendered with unprecedented texture fidelity 🔤 Reliable text & layout generates complex infographics timelines and multilingual slides with correct structure and legible text 🏆 Evaluated in 10k+ blind tests: now the strongest open-source image modelrivaling closed alternatives. 💡 Supports" [X Link](https://x.com/ModelScope2022/status/2006302672073507194) 2025-12-31T09:53Z [----] followers, [----] engagements "2026 is almost here Wishing you a very Happy New Year 🎉 What open-source models or product features are at the top of your wishlist for the coming year 🚀 Wed love to see if we can help make those wishes come true" [X Link](https://x.com/ModelScope2022/status/2006341719198163109) 2025-12-31T12:28Z [----] followers, [----] engagements "🌟Qwen-Image-2512 is now live on Model: Try: https://modelscope.ai/civision/imageGeneration/tab=default https://modelscope.ai/models/Qwen/Qwen-Image-2512/summary http://ModelScope.ai https://modelscope.ai/civision/imageGeneration/tab=default https://modelscope.ai/models/Qwen/Qwen-Image-2512/summary http://ModelScope.ai" [X Link](https://x.com/ModelScope2022/status/2006419353856782394) 2025-12-31T17:37Z [----] followers, [----] engagements "🚀 New on ModelScope: IQuest Research just dropped IQuest-Coder-V1 a new family of open-source code LLMs (7B to 40B) with 128K context GQA and two specialized variants: - Instruct: for coding assistance - Thinking: RL-tuned for agentic reasoning & error recovery - Plus: a novel Loop architecture (40B only) for efficient deployment via parameter sharing across [--] iterations. 🌟 All models support repo-level understanding via Code-Flow training learning from commit sequences and code evolution not just static snippets. 🔍 Benchmarks: - [----] on SWE-Bench Verified - [----] on LiveCodeBench v6" [X Link](https://x.com/ModelScope2022/status/2007712347373351418) 2026-01-04T07:14Z [----] followers, 10.9K engagements "Big news 🚀 now officially supports FREE LoRA training for Qwen-Image-2512 Z-Image-Turbo and Qwen-Image-Edit-2511. Time to build 🛠 🎨ModelScope Civision: https://modelscope.ai/civision/modelTraining http://ModelScope.ai https://modelscope.ai/civision/modelTraining http://ModelScope.ai https://modelscope.ai/civision/modelTraining http://ModelScope.ai https://modelscope.ai/civision/modelTraining http://ModelScope.ai" [X Link](https://x.com/ModelScope2022/status/2007787503802294735) 2026-01-04T12:13Z [----] followers, 21.5K engagements "Qwen-Image-2512-Fast is officially here ⚡ 🚀 Try the Demo on ModelScope now: Experience a massive 20x speedup generating high-quality images in just [--] to [--] steps Using CFG distillation weve achieved lightning-fast inference without compromising quality. https://modelscope.cn/studios/kelseye/Qwen-Image-2512-Turbo-LoRA-Demo/summary https://modelscope.cn/studios/kelseye/Qwen-Image-2512-Turbo-LoRA-Demo/summary https://modelscope.cn/studios/kelseye/Qwen-Image-2512-Turbo-LoRA-Demo/summary https://modelscope.cn/studios/kelseye/Qwen-Image-2512-Turbo-LoRA-Demo/summary" [X Link](https://x.com/ModelScope2022/status/2007792938248708207) 2026-01-04T12:35Z [----] followers, 30.5K engagements "@WolfyBlair Yes you'll get daily free quota (magicubes) and you can earn more by participating in community activities" [X Link](https://x.com/ModelScope2022/status/2008002207099293874) 2026-01-05T02:26Z [----] followers, [---] engagements "🚀 New on ModelScope: The Wuli Team just released a Turbo LoRA that makes Qwen-Image-2512 inference 20x faster⚡ without sacrificing quality. Key updates: ✅ 4-step turbo inference ✅ Enhanced texture & detail ✅ Native ComfyUI support Model: Demo: https://modelscope.cn/studios/kelseye/Qwen-Image-2512-Turbo-LoRA-Demo https://modelscope.ai/models/Wuli-Art/Qwen-Image-2512-Turbo-LoRA https://modelscope.ai/models/Wuli-Art/Qwen-Image-2512-Turbo-LoRA Qwen-Image-2512-Fast is officially here ⚡ 🚀 Try the Demo on ModelScope now: https://t.co/LRShwwbiNh Experience a massive 20x speedup generating" [X Link](https://x.com/ModelScope2022/status/2008020913309298809) 2026-01-05T03:41Z [----] followers, 11.3K engagements "🤖 Introducing InternVLA-A1 now fully open-sourced Many VLA models follow instructions well in static scenes but struggle in dynamic environments (conveyor belts rotating platforms multi-robot setups). Why They see the presentbut cant imagine the future. InternVLA-A1 solution: unify perception imagination and action in one model: ✅ Scene understanding: Image + text task parsing ✅ Task imagination: Predict future frames reason about dynamics ✅ Guided control: Execute actions steered by visual foresight Powered by InternData-A1 - Large-scale high-quality simulated dataset InternVLA-A1 stays" [X Link](https://x.com/ModelScope2022/status/2008137224575992238) 2026-01-05T11:23Z [----] followers, 38.2K engagements "🚀 Small model BIG agency Meet Youtu-LLM a native agentic 1.96B LLM that thinks plans and acts by itself no external frameworks needed. ✅ Outperforms 4B8B models on math coding & agent benchmarks ✅ Trained on 200B+ agentic reasoning trajectories ✅ 128K context open-weight & ready to run anywhere Think lightweight cant be smart Think again. 🤖 Model: 🔗 GitHub: 📄 Paper: https://www.modelscope.cn/papers/2512.24618 https://github.com/TencentCloudADP/youtu-tip/blob/master/youtu-llm https://modelscope.cn/collections/Tencent-YouTu-Research/Youtu-LLM https://www.modelscope.cn/papers/2512.24618" [X Link](https://x.com/ModelScope2022/status/2008447725600469060) 2026-01-06T07:57Z [----] followers, 11.8K engagements "🚀 TeleChat3-105B-A4.7B-Thinking is now open source A 105B sparse MoE model with fine-grained routing: - [---] experts only [--] activated per token (4.7B active params) - Trained end-to-end on domestic compute - Strong across code math agents writing check HumanEval-X (92.7%) & SWE-Bench (51%) It can even build a working shooter game 🕹 or a video site frontendfully autonomously. Also released: TeleChat3-36B-Thinking (dense 36B) optimized for reasoning & roleplay. ✅ Supports vLLM SGLang LLaMA-Factory ✅ Full Thinking mode with think./think reasoning traces ✅ Tuned for long-context and multi-turn" [X Link](https://x.com/ModelScope2022/status/2008499004158411187) 2026-01-06T11:20Z [----] followers, 16.8K engagements "The #EmbodiedAI data gap just got a lot smaller. 🚀 @AGIBOTofficial released the Genie Sim [---] datasetthe largest open-source sim dataset in the field 🤖 10000+ hours of high-quality sim demos 🎯 200+ tasks & 100k scenarios 📊 Multi-sensor: RGB-D Stereo & Kinematics Standardized in LeRobot format. Ready to train ✅ Dataset: ✅ GitHub: https://github.com/AgibotTech/genie_sim https://modelscope.cn/datasets/agibot_world/GenieSim3.0-Dataset https://github.com/AgibotTech/genie_sim https://modelscope.cn/datasets/agibot_world/GenieSim3.0-Dataset" [X Link](https://x.com/ModelScope2022/status/2008795887666147377) 2026-01-07T07:00Z [----] followers, [----] engagements "AI for all 🌍 Hosted by @alibaba_cloud the Alibaba Cloud AIGC Championship @ Milano Cortina [----] is officially live Create your Winter Olympics videos with the tools and resources powered by ModelScope as an official partner. ⛸❄ 🏆 Win Big 🎟 Milan [----] Tickets: A trip to the Winter Olympics 🏛 Legacy: Works archived in the Olympic Museum. 💰 Cash Prizes: Reward your creative storytelling. 🎬 Pro Tech: High-speed AI video via Wan [---]. Global Entrance Creating #ModelScope #AlibabaCloud #MilanoCortina2026 #Wan26 #Olympics https://modelscope.ai/civision/videoGeneration" [X Link](https://x.com/ModelScope2022/status/2008824590517321874) 2026-01-07T08:54Z [----] followers, [----] engagements "👏ComfyUI now supports Z-Image-Turbo LoRAs trained with ModelScope Civision Huge thanks to our amazing community members for the incredibly fast integration 👉 Everyone is welcome to train LoRAs for free on Civision and thanks to your valuable feedback training speed has been significantly improved 🚀 https://modelscope.ai/civision/modelTraining https://github.com/Comfy-Org/ComfyUI/pull/11805 https://github.com/Comfy-Org/ComfyUI/pull/11805 https://modelscope.ai/civision/modelTraining https://github.com/Comfy-Org/ComfyUI/pull/11805 https://github.com/Comfy-Org/ComfyUI/pull/11805" [X Link](https://x.com/ModelScope2022/status/2010551381233709197) 2026-01-12T03:16Z [----] followers, 11.9K engagements "GLM-Image is here Try it out and let us know what you think: https://x.com/ModelScope2022/status/2011262011997651194 🚀GLM-Image is now open-source a breakthrough in "Cognitive Generation." By combining Autoregressive models with Diffusion Decoders it finally solves the struggle of generating posters PPTs and knowledge-heavy infographics. 🌟 Highlights: ✅ #1 on CVTG-2K & LongText-Bench https://t.co/cF3z0VFvc9 https://x.com/ModelScope2022/status/2011262011997651194 🚀GLM-Image is now open-source a breakthrough in "Cognitive Generation." By combining Autoregressive models with Diffusion" [X Link](https://x.com/ModelScope2022/status/2011267263043944540) 2026-01-14T02:40Z [----] followers, [----] engagements "Step-Audio-R1.1 by @StepFun_ai just set a new SOTA on the Artificial Analysis Speech Reasoning leaderboard 🏆 It outperforms Grok Gemini and GPT-Realtime with a 96.4% accuracy rate. ✅ Native Audio Reasoning (End-to-End) ✅ Audio-native CoT (Chain of Thought) ✅ Real-time streaming inference ✅ FULLY OPEN SOURCE 🌟 Demo: 🤖 Model: https://modelscope.cn/models/stepfun-ai/Step-Audio-R1.1 https://modelscope.cn/studios/stepfun-ai/Step-Audio-R1 https://modelscope.cn/models/stepfun-ai/Step-Audio-R1.1 https://modelscope.cn/studios/stepfun-ai/Step-Audio-R1" [X Link](https://x.com/ModelScope2022/status/2011687986338136089) 2026-01-15T06:32Z [----] followers, 15.6K engagements "@SteveWarnerFL Hi would you mind sharing more about how you use it in your work Are you trying to modify a specific layer or add/remove one Wed love to better understand your situation" [X Link](https://x.com/ModelScope2022/status/2012050686171889776) 2026-01-16T06:34Z [----] followers, [--] engagements "Step3-VL-10B Demo: https://modelscope.cn/studios/stepfun-ai/step3-vl-10b https://modelscope.cn/studios/stepfun-ai/step3-vl-10b" [X Link](https://x.com/ModelScope2022/status/2013213460721475715) 2026-01-19T11:34Z [----] followers, [----] engagements "🚀 Introducing GLM-4.7-Flash the strongest 30B-class open model that balances performance & efficiency ✅ Top scores on AIME GPQA SWE-bench -Bench & more ✅ Optimized for agentic coding tool use and reasoning ✅ Runs locally with vLLM / SGLang / Transformers Perfect for developers who want Claude-level coding power at a fraction of the cost 💡 Try it now: 🤖 🔗 📄 🐙 https://github.com/zai-org/GLM-4.5 https://z.ai/blog/glm-4.7 https://chat.z.ai https://modelscope.cn/models/ZhipuAI/GLM-4.7-Flash https://github.com/zai-org/GLM-4.5 https://z.ai/blog/glm-4.7 https://chat.z.ai" [X Link](https://x.com/ModelScope2022/status/2013273444679590292) 2026-01-19T15:32Z [----] followers, 11.3K engagements "How reliable is your model's tool usage. REALLY 🧐 Inspired by @Kimi_Moonshot's K2 Vendor Verifier we are leveling up Agentic Evaluation with EvalScope. The standardizing Function Calling evaluation allows us to bridge the gap between "claims of support" and "RELIABLE execution" in the field. 🛠 📖 Docs: 💻 Code: 🔹 Standardized: Quantifies MoonshotAI K2 verification logic. 🔹 Comprehensive: Measures decision-making + JSON schema compliance. 🔹 Simple: Plug-and-play with your data or our benchmarks. Don't let your Agent go live without a proper check-up #LLM #AIAgents #FunctionCalling" [X Link](https://x.com/ModelScope2022/status/2014224614252253541) 2026-01-22T06:32Z [----] followers, [----] engagements "@YashasGunderia Hi you'll need to request to join the organization first and then you'll have access to the GPU resources. Here is the link: https://www.modelscope.ai/organization/xGPU-Explorers https://www.modelscope.ai/organization/xGPU-Explorers" [X Link](https://x.com/ModelScope2022/status/2014937070008205667) 2026-01-24T05:43Z [----] followers, [---] engagements "Weve prepared ModelScope and Tongyi swag for everyone attending the workshopready for a day full of great takeaways 🚀 Join Us in Seoul: AI Innovation Meets Creativity @AMD @Alibaba_Qwen @ModelScope2022 @alibaba_cloud AMDs AI Developer Meetup in Seoul (Dec 10) is filling FAST. As a key partner were bringing you the future of generative AI. 📅 Dec [--] 📍 Seoul 🎁 Free https://t.co/b5XcmEmbPf 🚀 Join Us in Seoul: AI Innovation Meets Creativity @AMD @Alibaba_Qwen @ModelScope2022 @alibaba_cloud AMDs AI Developer Meetup in Seoul (Dec 10) is filling FAST. As a key partner were bringing you the" [X Link](https://x.com/ModelScope2022/status/1998565458912620942) 2025-12-10T01:28Z [----] followers, [----] engagements "Qwen-Image-Layered-Control is now supported by ComfyUI Try it out: https://modelscope.ai/models/DiffSynth-Studio/Qwen-Image-Layered-Control/files DiffSynth-Studio has open-sourced Qwen-Image-Layered-Control 🚀 It enables precise text-controlled layer extractionperfect for poster decomposition and graphic design workflows. Control image layers like never before: ✅ Single-layer output via text prompts ✅ Native https://t.co/4QwOKqTFLs https://modelscope.ai/models/DiffSynth-Studio/Qwen-Image-Layered-Control/files DiffSynth-Studio has open-sourced Qwen-Image-Layered-Control 🚀 It enables precise" [X Link](https://x.com/ModelScope2022/status/2011733445324783764) 2026-01-15T09:33Z [----] followers, 22.9K engagements "Latency or intelligence Step-Audio-R1.1 ends the trade-off. Fei Tian Multimodal Large Model Researcher at StepFun introduces Mind-Paced Speakingenabling real-time voice AI that thinks while speaking. Fast responses. Deep reasoning. No compromise. https://x.com/i/broadcasts/1mrGmBabZavJy https://x.com/i/broadcasts/1mrGmBabZavJy" [X Link](https://x.com/ModelScope2022/status/2013219373948039628) 2026-01-19T11:57Z [----] followers, [----] engagements "Introducing AgentCPM-Report: An 8B on-device agent that rivals closed-source giants in deep research and professional report writing. The Performance: 🏆 Ranked #1 in Insight Depth across DeepResearch Bench Deep Consult and DeepResearch Gymoutperforming top-tier closed models. Highlights: ✅ 100% Private: Zero data leakage. Your sensitive data stays on your hardware. ✅ UltraRAG Stack: Efficient local indexing for private PDFs and docs. ✅ Production Ready: One-click Docker deployment. How does an 8B model hit SOTA 🧠 It breaks the "one-shot" generation bottleneck with a "Writing as Reasoning"" [X Link](https://x.com/ModelScope2022/status/2013586921046344058) 2026-01-20T12:18Z [----] followers, [----] engagements "🚀 Chroma [---] is here and its open From @flashlabsdotai : the worlds first open-source end-to-end real-time speech-to-speech dialogue model with personalized voice cloning. ✨ 150ms end-to-end latency 🧬 High-fidelity voice cloning from just seconds of audio 📈 Speaker similarity (SIM) = [-----] +10.96% over human baseline 🧠 Strong reasoning in only 4B params 🔓 Fully open weights + code A true open alternative to closed commercial systems and already optimized with @lmsysorgs SGLang for even faster inference 🔗 Paper: 🤖 Model: 💻 Code: Were excited to see what the community builds with" [X Link](https://x.com/ModelScope2022/status/2014006971855466640) 2026-01-21T16:07Z [----] followers, 14.3K engagements "🌾 Meet Sinong (): The first heavy-duty LLM suite for Agriculture 🚜 🧬 Expertise: Crop breeding animal medicine smart farming & ag-economics. 🤖 Models: 8B & 32B parameters 📚 Stack: 4B+ tokens (240k+ papers 8k+ books patents) 🧠 Tech: Native CoT reasoning + Multi-Agent RAG for zero-hallucination scientific retrieval. 🚀 Open Source: https://modelscope.cn/models/NAULLM/Sinong1.0-32B https://modelscope.cn/models/NAULLM/Sinong1.0-8B https://modelscope.cn/models/NAULLM/Sinong1.0-32B https://modelscope.cn/models/NAULLM/Sinong1.0-8B" [X Link](https://x.com/ModelScope2022/status/2014545364347715871) 2026-01-23T03:46Z [----] followers, [----] engagements "@noctus91 Hi there This isn't actually an error. To use API inference you just need to bind your Alibaba Cloud account (no extra cost involved)" [X Link](https://x.com/ModelScope2022/status/2014754313999155689) 2026-01-23T17:37Z [----] followers, [---] engagements "Meet LingBot-VA: The future of robot learning is visual. 🤖🎥 Its an autoregressive diffusion framework that predicts future video frames and decodes actions simultaneously. Instead of just reacting it reasons: "I expect the world to look like this next so I should move like that." Key Highlights: 🧠 Infinite Memory: Unlike memoryless models that get stuck in loops LingBot-VA handles "recurrent states" (like opening the same box twice) with ease by remembering the full history. 🍳 Generalist Skills: From folding clothes to making breakfast and high-precision screw pickingits a true" [X Link](https://x.com/ModelScope2022/status/2017067214919389569) 2026-01-30T02:47Z [----] followers, [----] engagements "ChenkinNoob-XL v0.2 is here A high-fidelity SDXL-to-Rectified Flow conversion delivering vivid color range and superior lighting adherence. ✅ Performance: Eliminates "base model greyness"; Superior contrast/dark scene rendering; High stability across wide CFG ranges (3-6); Optimal convergence at 20-28 steps. ✅ Ecosystem: Native ComfyUI workflow included; ReForge-ready with built-in Flow support; LoRA-compatible via Pochi.toml (Refresh scripts). ✅ Architecture: Rectified Flow (RF) linear ODE trajectory; Trained on 47M samples using 8xH20 nodes; Leverages Cosine Optimal Transport and Protected" [X Link](https://x.com/ModelScope2022/status/2019637177471173090) 2026-02-06T05:00Z [----] followers, [----] engagements "VibeThinker-1.5B is here 🚀 and it flips the bigger = smarter myth on its head. ✅ Just 1.5B params ✅ Trained via novel Spectrum-to-Signal Principle (SSP) ✅ Beats models 400x larger (e.g. 671B DeepSeek-R1) on hard math benchmarks (AIME24/25 HMMT25) ✅ Matches 456B MiniMax-M1 and ties Mistrals Magistral-Medium on LiveCodeBench v6 ✅ Post-training cost: $8K (vs. $290K+ for others) 💡 SSP first encourages divergent exploration of solution paths then uses RL to converge on optimal strategies smarter training not bigger models. 🌍 Open-sourced to empower researchers & smaller teams whove been priced" [X Link](https://x.com/ModelScope2022/status/1988955029064278216) 2025-11-13T13:00Z [----] followers, [----] engagements "Huge update for Z-Image-Turbo-Fun-Controlnet-Union [---] 🚀 Key technical upgrades: ✅ New 1.9GB Lite model for low-VRAM & natural blending ✅ Fixed mask leakage in inpainting ✅ Multi-resolution dataset refactor (up to 1536px) ✅ 8-step distillation for crisp blur-free Turbo gen No more bright spot artifacts. Just pure control. 🤖 Model: https://modelscope.ai/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union-2.1 https://modelscope.ai/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union-2.1" [X Link](https://x.com/ModelScope2022/status/2011356116119334957) 2026-01-14T08:34Z [----] followers, 20.5K engagements "Real-Qwen-Image-V2 is here 📸 Developed by wikeeyang this fine-tuned version of Qwen-Image-2512 is built for those who demand peak realism and production-grade quality. 🤖 Model: Key Highlights: ✨ Sharper details & enhanced realism ✨ Optimized for Asian facial aesthetics ✨ Superior text & image generation quality ✨ Highly compatible with LoRAs & custom workflows Tech Specs: ✅ CFG: [---] - [---] ✅ Steps: [--] - [--] ✅ Sampler: Euler / Simple (flexible) ✅ Model Shift: [---] - [---] https://www.modelscope.cn/models/wikeeyang/Real-Qwen-Image-V2 https://www.modelscope.cn/models/wikeeyang/Real-Qwen-Image-V2" [X Link](https://x.com/ModelScope2022/status/2012752195939168522) 2026-01-18T05:01Z [----] followers, 28.4K engagements "🚀 Meet STEP3-VL-10Bit delivers SOTA-level visual perception complex reasoning and human-aligned intelligenceredefining efficiency in open multimodal AI. ✅ Beats or matches models [----] larger (like GLM-4.6V Qwen3-VL even Gemini [---] Pro) ✅ Achieves SOTA on MMMU MathVision OCRBench ScreenSpot and more ✅ Trained on 1.2T tokens + 1400+ RL rounds (RLHF + RLVR) ✅ Supports PaCoRe: parallel collaborative reasoning (128K context) Despite its compact size it leads the 10B class in: STEM reasoning (94.43% on AIME [----] w/ PaCoRe) Visual perception (92.05 on MMBench) GUI understanding & OCR Spatial" [X Link](https://x.com/ModelScope2022/status/2013213019690639431) 2026-01-19T11:32Z [----] followers, 217.6K engagements "⚡🎨Tomorrow. This week is going to be intriguing" [X Link](https://x.com/ModelScope2022/status/2015613317088522594) 2026-01-26T02:30Z [----] followers, 51.8K engagements "DeepSeek-OCR-2 is now on ModelScope ✨ Meet DeepSeek-OCR-2 a vision-language model that reimagines document intelligence with human-like visual understanding. Key Highlights 🧠 Visual Causal Flow: Understands layout semantics and structurenot just pixels 📏 Dynamic Resolution: Processes documents at (0-6)768 + [----] with adaptive token generation ✍ Flexible Prompts: One model for multiple tasks: - Convert to markdown preserve tables & formatting - Parse the figure extract chart insights - Locate "text" pixel-level grounding - OCR this image 100+ language support ⚡ Production Ready:" [X Link](https://x.com/ModelScope2022/status/2016035065298616539) 2026-01-27T06:26Z [----] followers, [----] engagements "🚀 Meet Kimi K2.5 🌙 This is Kimis most intelligent and versatile model to date achieving SOTA performance across coding vision and agentic workflows. Model: Paper: Highlights: ✅ Native Multimodal Architecture: Seamlessly integrates vision & text. From screenshots to screen recordings K2.5 "sees" and reasons across images and videos to break the limits of text-only prompts. ✅ Agent Clusters: A world-first capability. K2.5 can autonomously spawn up to [---] "clones" to work in parallel on complex tasks slashing execution time by 4.5x. ✅ The Ultimate Office/Coding Assistant: * Office Pro: Masters" [X Link](https://x.com/ModelScope2022/status/2016043017610957071) 2026-01-27T06:58Z [----] followers, [----] engagements "The DiffSynth-Studio team also dropped Z-Image-i2L (Image-to-LoRA). 🎨✨ This model takes a single image as input and instantly outputs a custom LoRA tailored to that specific style or feature. Model: Demo: ✅ Zero Manual Effort: No captioning or training process required. ✅ One-Shot Magic: Get a style LoRA from just one reference image. ✅ Style Transfer: Seamlessly extract and apply unique features to new prompts. ✅ Lower Barrier: Personalized model creation is now accessible to everyone. https://modelscope.cn/studios/DiffSynth-Studio/Z-Image-i2L" [X Link](https://x.com/ModelScope2022/status/2016179961779126461) 2026-01-27T16:02Z [----] followers, [----] engagements "🚀 Meet LingBot-VLA: A pragmatic Vision-Language-Action model designed to bridge the gap between perception and execution in robotics. 🤖 ✅LingBot-VLA-4B: Lightweight & versatile. ✅LingBot-VLA-4B-Depth: Enhanced for high-precision spatial tasks. Highlights: 🧠 Powerful Core: Built on the Qwen2.5-VL-3B foundation mastering multi-tasking and dual-arm coordination across 9+ robot configs. 📈 Elite Performance: Outperforms competitors like [---] and GR00T in success rates (SR) on both GM-100 (Real-world) and RoboTwin [---] (Sim). 🔋 Hyper-Efficient: 1.52.8x faster training than existing VLA codebases" [X Link](https://x.com/ModelScope2022/status/2016209598471196691) 2026-01-27T18:00Z [----] followers, [----] engagements "Introducing LingBot-World: An open-source world simulator pushing the boundaries of video generation. 🚀 🌍 High-Fidelity: Realistic scientific & stylized. 🧠 Long-Term Memory: Minute-level consistency. ⚡ Real-Time: 1s latency at [--] FPS. 📜 Apache [---] Licensed. Model: Github: https://github.com/Robbyant/lingbot-world https://modelscope.cn/models/Robbyant/lingbot-world-base-cam https://github.com/Robbyant/lingbot-world https://modelscope.cn/models/Robbyant/lingbot-world-base-cam" [X Link](https://x.com/ModelScope2022/status/2016564553816088624) 2026-01-28T17:30Z [----] followers, 29.3K engagements "HunyuanImage-3.0 is here: A high-performance multimodal model for unified text-to-image generation and creative editing. 🚀🧵👇 🤖Model: 1/ The Model Suite 📦A versatile lineup for any workflow: ✅Base 3.0: The core T2I powerhouse. ✅3.0-Instruct: Advanced I2I and creative editing. ✅Instruct-Distil: Optimized for 8-step sampling and rapid deployment. 2/ Core Capabilities 🎨Beyond renderingits a creative partner. Features Unified Generation (T2I I2I and 3-image fusion) with built-in CoT reasoning and Prompt Enhancement to turn sparse ideas into professional descriptions. 3/ Technical Highlights" [X Link](https://x.com/ModelScope2022/status/2016841562274066648) 2026-01-29T11:51Z [----] followers, [----] engagements "PaddleOCR-VL-1.5 is open-sourced A 0.9B parameter multimodal model that masters the "curved & distorted" document challenge. 🚀 🔗 ModelScope: 🔗 Demo: Key Technical Breakthroughs: ✅Polygon-based Localization: No more rigid rectangles. It natively supports irregular boxes to fit warped tilted or screen-captured text perfectly. 📐 ✅Seal & Stamp Recognition: New dedicated capability to handle complex official document verification. 💮 ✅Cross-Page Logic: Automatically merges tables and identifies headings across physical page breakscrucial for full-document semantics. ✅Performance Giant:" [X Link](https://x.com/ModelScope2022/status/2017113730866434213) 2026-01-30T05:52Z [----] followers, [----] engagements "Tencent YouTu Research open-sourcedYoutu-VL-4B-Instruct a compact VLM redefining visual perception viaVLUAS. 👁⚡ 🌟Highlights: ✅All-in-One Vision: SOTA performance in detection segmentation depth estimation and pose estimationwithouttask-specific heads. ✅OCR & Reasoning: Exceptional at complex document parsing and multimodal math reasoning. ✅GUI Agent Ready: Optimized for world-understanding and interface navigation. ✅Efficiency: 4B parameters make it perfect for edge deployment and fast inference. 🔧Performance: Outperforms many larger models inOmniDocBenchandVision-Centrictasks while" [X Link](https://x.com/ModelScope2022/status/2017191058262700328) 2026-01-30T11:00Z [----] followers, [----] engagements "Ready to make this February legendary 🚀✨The Qwen-Image LoRA Training Competition is officially LIVE Weve teamed up with @Ali_TongyiLab and @Alibaba_Qwen to bring you some absolutely grand prizes:📱 iPhone [--] Pro Max (512GB) 🎮 Sony PS5 Pro 🛍 $800 Shopping Cards (or equivalent prizes) Stop scrolling and start training 🛠🎨Join here: #HappyQwensday #QwenImageLoRA https://modelscope.ai/active/qwenimagelora https://modelscope.ai/active/qwenimagelora" [X Link](https://x.com/ModelScope2022/status/2017939891846512907) 2026-02-01T12:35Z [----] followers, [----] engagements "Beyond-Reality-Z-Image [---] is here.🚀 Highlights: ✅Texture SOTA: Optimized high-frequency details for what might be the best out-of-the-box skin and fabric textures in the Z-model family. ✅Cinematic Lighting: Re-engineered color balance to achieve that authentic "film look" without heavy prompting. ✅Speed x Versatility: Generates at Turbo speeds (10-15 steps) but supports broader artistic styles and 2D/Anime aesthetics than the official release. ✅Efficiency: Optimized for low CFG (1.0) making it highly stable and responsive to LoRA integration. Available now on ModelScope: Recommended Stack:" [X Link](https://x.com/ModelScope2022/status/2018309394719805444) 2026-02-02T13:03Z [----] followers, 16.4K engagements "2/3 🛠Four specialist models were released in just one week each mastering a specific real-world challenge: ✅ DeepSeek-OCR2: Uses "Visual Causal Flow" to handle reading order in complex multi-column layouts. ✅ GLM-OCR (Z-org): An efficiency powerhouse using Multi-Token Prediction (MTP) to reduce costs by 90% while excelling at formulas. ✅ PaddleOCR-VL-1.5 (PaddlePaddle): Built for "Real5" messy conditionsperfect for curved pages and camera glare. ✅ Youtu-Parsing (Tencent): Delivers 22x faster inference and converts flowcharts directly into Mermaid code." [X Link](https://x.com/ModelScope2022/status/2018615134479057183) 2026-02-03T09:18Z [----] followers, [----] engagements "Here they come They are now at or alumni of these elite AI labs: Alibaba Tongyi Meta (FAIR/Superintelligence) Google DeepMind Tencent AI. Academia: SJTU OSU UChicago NUS WashU. https://x.com/i/broadcasts/1ZkJzZRdQNoJv https://x.com/i/broadcasts/1ZkJzZRdQNoJv" [X Link](https://x.com/ModelScope2022/status/2019288799780770192) 2026-02-05T05:55Z [----] followers, [----] engagements "@pranaysuyash http://ModelScope.ai http://ModelScope.ai" [X Link](https://x.com/ModelScope2022/status/2019390286355656859) 2026-02-05T12:38Z [----] followers, [---] engagements "1/4🧵 MS-Agent is Rebuilding A massive functional update to ModelScopes agent frameworkintroducing autonomous skill scaling and industrial-grade toolkits. ✅ Agent Skills: Skill Discovery via local/Hub repo_id matching; ✅ Multi-skill Collaboration via DAG-based planning; ✅ Progressive Analysis to optimize context window for small models. ✅ Robust Executor: Secure Docker/Sandbox execution; autonomous self-reflection and bug-fix cycles. Github: https://github.com/modelscope/ms-agent/tree/main/ms_agent/skill https://github.com/modelscope/ms-agent/tree/main/ms_agent/skill" [X Link](https://x.com/ModelScope2022/status/2019600619443744930) 2026-02-06T02:34Z [----] followers, [---] engagements "2/4 💻 Code Genesis is Now Live A SOTA end-to-end synthesis engine specializing in zero-to-one development over standard "vibe-coding." ✅ Capabilities: Industrial-grade Code Genesis for full-stack synthesis; Language Server integration for stability and clean code aesthetics. ✅ Technical Edge: Superior benchmarks in end-to-end logic flow; minimalist architecture for rapid project scaffolding. Repo: Tech Report: https://github.com/modelscope/ms-agent/tree/main/projects/code_genesis/TECH_REPORT.md https://github.com/modelscope/ms-agent/tree/main/projects/code_genesis" [X Link](https://x.com/ModelScope2022/status/2019600621608022159) 2026-02-06T02:34Z [----] followers, [---] engagements "🎬 Singularity Cinema is Breaking Benchmarks The GLCC-winning workshop for high-density long-form automated video production. ✅ Generation: LLM-driven storyboards; Manim/Remotion dual-engine animation; 5min+ cinematic output for Science Econ and History. ✅ Model Fusion: Seamlessly integrates Qwen-Image Sora edge-tts and nano-banana-pro; transforms static docs into dynamic media. Demo: https://github.com/modelscope/ms-agent/tree/main/projects/singularity_cinema https://github.com/modelscope/ms-agent/tree/main/projects/singularity_cinema" [X Link](https://x.com/ModelScope2022/status/2019600623738765722) 2026-02-06T02:34Z [----] followers, [---] engagements "4/4 🌍 ms-agent WebUI is Live A self-bootstrapped interface generated entirely by the Code Genesis engine. ✅ Usage: pip install ms-agent -U - ms-agent ui; instant GUI access for agent orchestration. ✅ Roadmap: DeepResearch capability launching soonSOTA autonomous retrieval and synthesis incoming. Full Framework: https://github.com/modelscope/ms-agent https://github.com/modelscope/ms-agent" [X Link](https://x.com/ModelScope2022/status/2019600625835864392) 2026-02-06T02:34Z [----] followers, [---] engagements "Thinking Machines Lab proved On-Policy Distillation slashes LLM training costs by 10x and we show you how to reproduce their research. Invest [--] minutes in this guideas we unpack the theory tech details experiment results and code to instantly transform your fine-tuning budget📚 Related Resources👇): ✅ Slash training compute by 10X. ✅ Achieve robust RL performance with zero forgetting. ✅ Get the ready-to-use ms-SWIFT + vLLM code for deployment. Related Resources - TML Blog: - (Open-source implementation for reproducing On-Policy Distillation) - On-Policy Distillation Documentation: - Example" [X Link](https://x.com/ModelScope2022/status/1983159127934488946) 2025-10-28T13:09Z [----] followers, [----] engagements "Stepfun open-sourcedStep-3.5-Flash a powerhouse model specifically architected for high-speed reasoning and complex Agentic workflows. 🚀 Model: Key Technical Specs: ✅Sparse MoE Architecture: 196B total params but only11B activeper token. SOTA efficiency. ✅MTP-3 (Multi-Token Prediction): It predicts3 tokens at once hitting a blistering350 TPSfor code-heavy tasks. ⚡ ✅Hybrid Attention (SWA + Full): A 3:1 mix that masters256K contextwindows while keeping compute costs low. ✅Parallel Thinking: Massively boosted performance for multi-step reasoning and deep search. Why Devs should care: - Built" [X Link](https://x.com/ModelScope2022/status/2018215995740164168) 2026-02-02T06:52Z [----] followers, 34K engagements "How do Agents "Grow from Experience" A dialogue on Key Issues and Technical Paths of Self-Evolution AI AMA Deep Roundtable Vol.1 [----] was about Reasoning. [----] is the "Mirror Moment"Agents are shifting from static tools to digital lifeforms that learn iterate and evolve from every failure. 🚀 https://twitter.com/i/web/status/2018712297829732601 https://twitter.com/i/web/status/2018712297829732601" [X Link](https://x.com/ModelScope2022/status/2018712297829732601) 2026-02-03T15:44Z [----] followers, [----] engagements "Meetthe1stAuthorsbehindthemostinfluentialpapers:earlyexperienceDREAMGYMAgentEvolverSPICER-ZeroEvo-MemoryandSEAgent. TheyarenowatoralumnioftheseeliteAIlabs: 🔹AlibabaTongyiLabMetaFAIRMetaSuperintelligenceLabsGoogleDeepMindTencentAISeattleLab 🔹SJTUOSUUChicagoNUSWashU 6leadresearchers1stage.Unrivaledtechnicaldepth.🧠 https://twitter.com/i/web/status/2018712300744847796 https://twitter.com/i/web/status/2018712300744847796" [X Link](https://x.com/ModelScope2022/status/2018712300744847796) 2026-02-03T15:44Z [----] followers, [---] engagements "MiniCPM-o [---] brings true "Omni" capabilities (Vision + Audio + Text) into a single 9B end-to-end architecture (Qwen3-8B + SigLip2 + Whisper + CosyVoice2).🚀 ✅ Vision SOTA: OpenCompass score of 77.6surpassing GPT-4o and Gemini [---] Pro in VLM tasks with just 9B params. ✅ Real-time Full Duplex: It "sees" "hears" and "speaks" simultaneously. Supports active interaction (it can interrupt or initiate conversation based on live video). ✅ World-Class OCR: Beats DeepSeek-OCR [--] and GPT-4o in document parsing (OmniDocBench leader). ✅ Voice Cloning: Clone any voice with just a short reference" [X Link](https://x.com/ModelScope2022/status/2018925461649396202) 2026-02-04T05:51Z [----] followers, [----] engagements "🚀 Meet Intern-S1-Pro: A massive 1T parameter MoE model for Multimodal Science Reasoning ✅ [---] Experts (22B active) ✅ SOTA in AI4Science (Chemistry Materials Bio) ✅ FoPE + Time-series modeling (up to [--] points) ✅ Native "Thinking Mode" support Open-source science just leveled up. 🧪💻 Model: Github: https://github.com/InternLM/Intern-S1 https://www.modelscope.cn/models/Shanghai_AI_Laboratory/Intern-S1-Pro https://github.com/InternLM/Intern-S1 https://www.modelscope.cn/models/Shanghai_AI_Laboratory/Intern-S1-Pro" [X Link](https://x.com/ModelScope2022/status/2019068430041919653) 2026-02-04T15:20Z [----] followers, [----] engagements "From a local designer in Chengdu to a global AI sensation with 117k+ downloadsDaniels journey shows whats possible when open-source meets imagination. Now its your turn. Join the Qwen-Image LoRA Competition and start building your own legacy. The next top creator could be you. https://modelscope.ai/active/qwenimagelora LoRA Training Contest: Win iPhone/PS5 and other prizes Gaussian splash LoRA creation process shared. Competition information: https://t.co/MEEZHQcr83 Training address: https://t.co/uxwtMlJzoP @Ali_TongyiLab @ModelScope2022 #HappyQwensday #QwenImageLora https://t.co/5igMVZUO5i" [X Link](https://x.com/ModelScope2022/status/2019612106086314106) 2026-02-06T03:20Z [----] followers, [----] engagements "StepFun open-sourced Step-3.5-Flashfeaturing powerful reasoning and Agent intelligence. Speeds up to [---] TPS Competes with closed-source models in Agent and Math tasks Excels at complex long-chain reasoning. Dont miss out https://x.com/i/broadcasts/1OyKAjZgZPaGb https://x.com/i/broadcasts/1OyKAjZgZPaGb" [X Link](https://x.com/ModelScope2022/status/2019744317792018474) 2026-02-06T12:05Z [----] followers, [----] engagements "YOLO26 is here🚀 The most advanced and deployment-ready vision model from @ultralytics redefining real-time edge computing performance. ✅ Performance: 43% CPU inference speed boost vs YOLO11; Real-time performance on low-power hardware; [--] scalable sizes (n/s/m/l/x) for optimized speed-accuracy trade-offs. ✅ Ecosystem: Unified framework for Detection Segmentation Pose OBB and Tracking; Simplified export for seamless hardware integration. ✅ Logic/Architecture: Removed DFL for enhanced hardware compatibility; End-to-end NMS-free inference; ProgLoss + STAL for small target precision; MuSGD" [X Link](https://x.com/ModelScope2022/status/2020777189575467356) 2026-02-09T08:30Z [----] followers, [----] engagements "2/4 Key Technical Highlights: - Architecture: Native single-stream design integrating generation editing and segmentation. - Audio: First model to unify Speech Audio and Music generation via 12.5Hz continuous representation. - Video: Time-Interleaved VideoRoPE for enhanced temporal modeling. https://twitter.com/i/web/status/2021580867009671592 https://twitter.com/i/web/status/2021580867009671592" [X Link](https://x.com/ModelScope2022/status/2021580867009671592) 2026-02-11T13:43Z [----] followers, [---] engagements "🐿 Meet Sirchmunk: The agentic search engine that operates directly on your raw data. Turn files into evolving intelligence in real-time Build information retrieval pipelines that require ZERO vector databases ZERO pre-indexing and ZERO chunking: 🎯 Embedding-Free Retrieval Drop files in and search instantly. No ETL pipelines no vector silos and no setup hours. 🎲 Monte Carlo Evidence Sampling It doesn't "read" full docsit explores focuses and synthesizes. 500-page manual Only the most relevant snippets hit the LLM. 🧠 Self-Evolving Knowledge Every search builds reusable knowledge clusters." [X Link](https://x.com/ModelScope2022/status/2022304582207238586) 2026-02-13T13:39Z [----] followers, [---] engagements "LongCat-Flash-LiteisNow Live Meituan's 68.5B MoE model pioneering "Embedding Expansion" for leading Agent and Coding performance. ✅Performance: 2.9B4.5B activated params; 500-700 tokens/s throughput; 256K context via YARN; SOTA SWE-Bench (54.4%) and -Bench scores. ✅Ecosystem: Optimized SGLang-FluentLLM engine with custom CUDA kernels; Open API with 50M free daily tokens. ✅Architecture: N-gram embedding layer for localized semantic precision; Hash-collision mitigation via sub-table decomposition; Triple-step Speculative Decoding synergy. 🤖 🔧 📄 Tech" [X Link](https://x.com/ModelScope2022/status/2020383958819832285) 2026-02-08T06:27Z [----] followers, [----] engagements "🚀 GLM-4.6V is live @Zai_org 's new open-source multimodal models just dropped on ModelScope and we provide a FREE API for immediate access❗ GLM-4.6V (106B) for cloud & high-performance workloads GLM-4.6V-Flash (9B) lightweight fast great for local inference 🔥 Key dev perks: ✅ Native multimodal tool calling pass images/docs directly as function args no OCR detour ✅ 128K context handles 150-page docs or hour-long videos in one go ✅ Visual Action pipeline powers real multimodal agents (e.g. find this outfit online returns structured shopping list) ✅ 50% cheaper than GLM-4.5V $1/million input" [X Link](https://x.com/ModelScope2022/status/1997995482459865290) 2025-12-08T11:43Z [----] followers, [----] engagements "Nanbeige just dropped Nanbeige4.1-3B a tiny powerhouse that outclasses models 10x its size.🚀 ✅ 256k Context Window ✅ "Deep Search" Agent capabilities built-in ✅ Two-stage Code RL (Correctness ➡ Efficiency) ✅ Beats Qwen3-32B on key benchmarks The era of high-performance low-cost local deployment is here. 🤖: https://modelscope.cn/models/nanbeige/Nanbeige4.1-3B https://modelscope.cn/models/nanbeige/Nanbeige4.1-3B" [X Link](https://x.com/ModelScope2022/status/2022629650178941435) 2026-02-14T11:11Z [----] followers, 10.9K engagements "Qwen3-ASR & Qwen3-ForcedAligner are now open-source 🎙🚀 High-performance streaming-ready speech AI for [--] languages and dialects. 🔗: Demo: 🤖The Models ✅ ASR-1.7B: Open-source SOTA. Outperforms Whisper-v3; rivals GPT-4o & Gemini. ✅ ASR-0.6B: Insane efficiency92ms latency; transcribes 2000s of audio in 1s. ✅ ForcedAligner-0.6B: Superior timestamp accuracy across [--] languages. 🌟Key Features ✅Universal: Supports [--] languages + [--] Chinese dialects. ✅Robust: Expertly handles singing heavy BGM and low SNR. ✅Unified: Single model for both streaming & offline (up to 20min audio). ⚡Performance &" [X Link](https://x.com/ModelScope2022/status/2016857817597301174) 2026-01-29T12:55Z [----] followers, [----] engagements "🚀 ACE-Step v1.5 is out: an open-source music foundation model that runs locally on consumer GPUs (4GB VRAM) and generates full songs in 2s (A100) or 10s (RTX 3090). ✅ Beats most commercial models in quality ✅ Train a personalized LoRA from just a few tracks ✅ Built on a novel LM+DiT hybrid architecture with internal RLno external reward bias ✅ Supports 50+ languages + editing (covers repainting vocalBGM) 🤖 https://modelscope.cn/models/ACE-Step/Ace-Step1.5 https://modelscope.cn/models/ACE-Step/Ace-Step1.5" [X Link](https://x.com/ModelScope2022/status/2018977724048777344) 2026-02-04T09:19Z [----] followers, [----] engagements "Introducing Singularity Cinema (MS-Agent framework)🚀an LLM-powered tool that transforms documents or prompts into storyboards generating stunning 5-minute+ videos via its Manim/Remotion dual-engine. By fusing nano-banana-pro🍌 Qwen-Image🎨 Sora and edge-tts it delivers high-quality information-dense contentperfect for science economics and history creators looking to scale their social media presence. 🔗: 🎦Prompt: Convert xx/deploy.md into a short video in a blue-themed style making sure to use the important images from the document. The short video must be in English. 🎬Output Video👇" [X Link](https://x.com/ModelScope2022/status/2019443081461002520) 2026-02-05T16:08Z [----] followers, [----] engagements "Meet X-Coder 🚀IIGroup drops a powerhouse suite for code LLMs ✅ X-Coder-SFT-376k: Massive synthetic dataset via GPT-o3-mini/DeepSeek-R1. ✅ X-Coder-RL-40k: Targeted RL alignment for logic & verification. ✅ X-Coder-8B Models: SOTA reasoning built on Qwen3. Experience the power of fully synthetic code evolution. 💻🔥 📊 Models & Dataset: #LLM #OpenCode #Qwen3 #DeepSeek #Coding https://modelscope.cn/models/IIGroup/X-Coder-SFT-Qwen3-8B https://modelscope.cn/datasets/IIGroup/X-Coder-RL-40k https://modelscope.cn/datasets/IIGroup/X-Coder-SFT-376k" [X Link](https://x.com/ModelScope2022/status/2020745735059443936) 2026-02-09T06:25Z [----] followers, [----] engagements "🚀Introducing Z-Image the full non-distilled version of Z-Image for high-quality customizable image generation. ✅ Model: ✅ Demo: ✨ Capabilities: Photorealistic + diverse art styles (anime illustration digital art) Full CFG & precise negative prompts for reliable prompt following Multi-subject generation with distinct identitiesno "clone faces" Community-friendly: ready for LoRA ControlNet and other fine-tuning workflows 🔥 Highlights: Full architecture (vs. distilled Turbo): better diversity & training stability ideal for customization High output variance across seeds = less homogenization" [X Link](https://x.com/ModelScope2022/status/2016178781522890948) 2026-01-27T15:57Z [----] followers, 59.4K engagements "How 0.9B Models Are Outperforming the Giants: A Comparison of Four New Chinese VLM-OCR Models 1/3 🏆 Early [----] has brought a massive surge of open-source models from DeepSeek Z-org PaddlePaddle and Tencent YouTu. A tiny 0.9B parameter model GLM-OCR topped the global OmniDocBench v1.5 leaderboard outscoring giants like Gemini-3 Pro and Qwen3-VL. We are moving away from rigid "detect and recognize" pipelines toward Vision-Language Models (VLMs) that actually "read" and understand document structures like humans do. https://twitter.com/i/web/status/2018615132134551565" [X Link](https://x.com/ModelScope2022/status/2018615132134551565) 2026-02-03T09:18Z [----] followers, 30K engagements "Z-Image-Fun-Controlnet-Union & Lora-Distill are now open-source 🚀Expanding the Z-Image ecosystem with ultra-fast inference and high-precision spatial control. ✅ Performance: Achieve 8-step inference (CFG: 1.0) via Distill LoRA and professional-grade multi-condition control (Canny Depth Pose Inpaint) across 15+ layer blocks. ✅ Ecosystem: Full cross-compatibility between Distill LoRA and Union ControlNet; includes "Lite" versions for low-VRAM edge devices and seamless natural blending. ✅ Unified Control: Support for 7+ spatial conditions in a single model with adjustable control_context_scale" [X Link](https://x.com/ModelScope2022/status/2019339552650092830) 2026-02-05T09:17Z [----] followers, 10.9K engagements "Z-Image-Fun-Lora-Distill is here A high-efficiency Distill LoRA enabling CFG-free high-fidelity generation for Z-Image in just 4-8 steps. ✅ 4-step and 8-step variants; CFG-free inference (CFG=1.0); [----] update delivers superior skin texture and color consistency compared to previous iterations. ✅ Native ComfyUI weights available; full integration with the VideoX-Fun framework; compatible with Z-Image ControlNet Union [---] and third-party LoRAs. ✅ Dual-distillation of both sampling steps and CFG; trained from scratch without Z-Image-Turbo weights; optimized for rapid generation and derivative" [X Link](https://x.com/ModelScope2022/status/2021412342723207600) 2026-02-11T02:33Z [----] followers, [----] engagements "🚀1/4 Introducing Ming-flash-omni-2.0: An open-source Omni-MLLM breakthrough. Based on the Ling-2.0 MoE architecture (100B total / 6B active) it unifies vision speech and image editing into a single model matching specialist performance across all modalities. 🧵👇 🤖 ModelScope: 📦 GitHub: https://github.com/inclusionAI/Ming https://www.modelscope.cn/models/inclusionAI/Ming-flash-omni-2.0 https://github.com/inclusionAI/Ming https://www.modelscope.cn/models/inclusionAI/Ming-flash-omni-2.0" [X Link](https://x.com/ModelScope2022/status/2021580863277084888) 2026-02-11T13:43Z [----] followers, [----] engagements "🚀 GLM-5 is now open-sourceand ranked #4 globally #1 among open models on Artificial Analysis Beyond "vibe coding" it delivers Agentic Engineering: autonomously architects debugs and ships production systems. Coding performance rivals Claude Opus [---] (SWE-bench: [----] 🔓 SOTA) while dominating long-horizon agent tasks (BrowseComp MCP-Atlas 🔝). 744B params MIT License 👉 ModelScope: https://modelscope.cn/models/ZhipuAI/GLM-5 https://modelscope.cn/models/ZhipuAI/GLM-5" [X Link](https://x.com/ModelScope2022/status/2021792323227820181) 2026-02-12T03:43Z [----] followers, [----] engagements "Love seeing what the MiniMax team shipped with M2.5 🚀 Real-world productivity strong agentic workflows and impressive efficiency numbers this is serious engineering. SOTA across coding search and tool use while getting faster is a big deal. 🔥 Excited to share that the model will be coming to ModelScope soon for the open-source community. Stay tuned 🌍✨ Introducing M2.5 an open-source frontier model designed for real-world productivity. - SOTA performance at coding (SWE-Bench Verified 80.2%) search (BrowseComp 76.3%) agentic tool-calling (BFCL 76.8%) & office work. - Optimized for efficient" [X Link](https://x.com/ModelScope2022/status/2021994964125987267) 2026-02-12T17:09Z [----] followers, [----] engagements "🚀 Ring-2.5-1T is open-source The first hybrid linear 1T thinking model. 📊 Gold Tier reasoning IMO25 (35/42) & CMO25 (105/126) scores. 🛠 Agent-native compatible with Claude Code OpenClaw and SGLang. ⚡Hybrid Linear breakthrough 10x lower memory & 3x throughput for long context. 🧠 1T MoE structure with 63B active parameters (1:7 MLA + Lightning Linear). 💻 Long-Horizon mastery autonomously built a 32-bit OS from scratch in [--] hours. Model: GitHub: https://github.com/inclusionAI/Ring-V2.5 https://modelscope.cn/models/inclusionAI/Ring-2.5-1T https://github.com/inclusionAI/Ring-V2.5" [X Link](https://x.com/ModelScope2022/status/2022223579421839523) 2026-02-13T08:17Z [----] followers, 12.9K engagements "🚀 FantasyWorld is now open source Recently accepted to ICLR [----] and ranking #1 on the WorldScore leaderboard this framework offers a unified approach to video and 3D generation. ✨Key Technical Highlights: ✅Direct 3D Inference: Input an image and camera trajectory to get frame-consistent video dense depth maps and point clouds instantly. ✅Dual-Branch IRG Architecture: An "Imagination" branch handles visual appearance while a "Geometry" branch ensures spatial accuracy. ✅Wan [---] Integration: The released 14B models leverage the latest Wan-DiT architectures for superior temporal consistency and" [X Link](https://x.com/ModelScope2022/status/2022264617871757514) 2026-02-13T11:00Z [----] followers, [----] engagements "🐿 Meet Sirchmunk: The agentic search engine that operates directly on your raw data. Turn files into evolving intelligence in real-time Build information retrieval pipelines that require ZERO vector databases ZERO pre-indexing and ZERO chunking: 🎯 Embedding-Free Retrieval Drop files in and search instantly. No ETL pipelines no vector silos and no setup hours. 🎲 Monte Carlo Evidence Sampling It doesn't "read" full docsit explores focuses and synthesizes. 500-page manual Only the most relevant snippets hit the LLM. 🧠 Self-Evolving Knowledge Every search builds reusable knowledge clusters." [X Link](https://x.com/ModelScope2022/status/2022370189157831078) 2026-02-13T18:00Z [----] followers, [----] engagements "Introducing FireRed-Image-Edit-1.0 from FireRedTeam 🚀 Its officially the new SOTA for general image editing. ✅ Better than Closed-Source: Outperforms Nano-Banana & Seedream4.0 on GEdit benchmarks. ✅ Native Evolution: Built from T2I foundations not just a "patch" on existing models. ✅ Style Mastery: Scored a record-breaking 4.97/5.0 in style transfer. ✅ High-Fidelity Text: Keeps original font styles perfectly. ✅ Virtual Try-on: Native support for multi-image joint editing. ✅ Bilingual: Native support for both English & Chinese prompts. Apache [---] license. Local deployment ready. 🤖 Model: 🎠" [X Link](https://x.com/ModelScope2022/status/2022687312442704320) 2026-02-14T15:00Z [----] followers, 53.9K engagements Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing
@MaaSAI42 ModelScope
ModelScope posts on X about ai, native, agentic, inference the most. They currently have [-----] followers and [---] posts still getting attention that total [------] engagements in the last [--] hours.
Engagements: [------] #
- [--] Week [------] -12%
- [--] Month [-------] +72%
- [--] Year [---------] +210,270%
Mentions: [--] #
- [--] Year [--] +9,400%
Followers: [-----] #
- [--] Week [-----] +4.50%
- [--] Month [-----] +52%
CreatorRank: [-------] #
Social Influence
Social category influence technology brands travel destinations stocks finance products gaming
Social topic influence ai, native #166, agentic #244, inference #72, llm, model, realtime, the first, math, strong
Top accounts mentioned or mentioned by @alibabaqwen @furaidosu @merjicai @alitongyilab @meituanlongcat @stepfunai @dx8152 @kimimoonshot @nhldesktop @zhihufrontier @tencenthunyuan @alibabacloud @pranaysuyash @ali_tongyilab @barrakali @1328 @github @baiduinc @isaifai @crazytoolman
Top assets mentioned Robot Consulting Co., Ltd. (LAWR) Alphabet Inc Class A (GOOGL)
Top Social Posts
Top posts by engagements in the last [--] hours
"@Alibaba_Qwen @github yes we are still here😉🙌🙌"
X Link 2024-09-05T07:18Z [--] followers, [---] engagements
"🚀 We have launched nano-banana Studio You can try it on ModelScope for FREE🔗Link in the first comment 🔥 Here are [--] tips for beginners to make prompts clear stable and controllable👇 [--] Lock what must not change Tell the model what to preserve: character pose expression colors. Example: Keep the main character unchanged. [--] Start with verbs Be direct: remove / add / replace / turn into. Avoid vague phrasing like adjust a bit. [--] Be ultra-specific Describe object + color + light + material + style. Example: Change the background to pure white no shadows. [--] Layer your prompt Structure = Subject"
X Link 2025-08-27T07:56Z [---] followers, [---] engagements
"🚀 Exciting update from @Kimi_Moonshot Kimi K2-0905 brings enhanced coding 256k context and better agent integration. Check out the impressive benchmarks vs Claude Sonnet [--] and try the turbo API for 100% tool-call accuracy. 🔗 https://modelscope.cn/models/moonshotai/Kimi-K2-Instruct-0905 Kimi K2-0905 update 🚀 - Enhanced coding capabilities esp. front-end & tool-calling - Context length extended to 256k tokens - Improved integration with various agent scaffolds (e.g. Claude Code Roo Code etc) 🔗 Weights & code: https://t.co/83sQekosr9 💬 Chat with new Kimi https://t.co/mkOuBMwzpw"
X Link 2025-09-05T03:41Z [---] followers, [---] engagements
"🎨✨ Say hello to HunyuanImage-2.1 your new open-source AI art powerhouse: - 2K resolution at lightning speed - Multi-language prompts (English/Chinese) - Flexible aspect ratios (1:1 16:9 9:16 4:3 3:4 3:2 2:3) - Smart PromptEnhancer for effortless detail & style - Quality on par with Seedream3.0 and Qwen-Image 🔗 Check it out: #hunyuanimage #AIart https://modelscope.cn/models/Tencent-Hunyuan/HunyuanImage-2.1/summary https://modelscope.cn/models/Tencent-Hunyuan/HunyuanImage-2.1/summary"
X Link 2025-09-09T09:09Z [---] followers, [---] engagements
"🚀Huge congrats to @Alibaba_Qwen Qwen3-Next is seriously innovative. As community members put it: Qwen just set the new baseline. Might be the first shot in the race for hyper-complex hybrid-arch models. Lets break down what it brings: ⚡ Smaller model size same capability 💰 [--] cheaper GPU hours 📜 [--] longer context 🥊 Beats Gemini-2.5-Flash-Thinking Full thread"
X Link 2025-09-12T08:42Z [---] followers, [---] engagements
"1 Hybrid Architecture - 75% of layers use efficient linear attention (Gated DeltaNet) - 25% keep enhanced standard attention balances long-context efficiency & recall fidelity Enhancements (standard-attention layers only): - Output gating to curb low-rank bottlenecks - Head dimension upsized: [---] [---] - RoPE applied only to first 25% of positions per head boosts length extrapolation"
X Link 2025-09-12T08:45Z [---] followers, [--] engagements
"2 Extreme Sparse MoE - 80B total parameters but only 3B activated per inference - Architecture: [---] experts + [--] routing experts + [--] shared expert Global load balancing decouples parameter scale from inference cost max efficiency without losing performance"
X Link 2025-09-12T08:47Z [---] followers, [--] engagements
"4 Native Multi-Token Prediction (MTP) - Provides an MTP module with high Speculative Decoding acceptance rates - Enhances backbones overall performance - Optimizes multi-step inference via inference-consistent multi-step training further improves Speculative Decoding in practical scenarios"
X Link 2025-09-12T08:49Z [---] followers, [--] engagements
"@Baidu_Inc has launched ERNIE X1.1 claiming it outperforms DeepSeek-R1-0528 and matches GPT-5 and Gemini [---] Pro in performance. Early users report noticeable improvements especially in reasoning. Worth a trycheck it out and test it in different scenarios ERNIE X1.1 just launched Unveiled at Wave Summit [----] our latest reasoning model reduces hallucinations improves instruction following and delivers strong agentic capabilities. ✅ Achieves significant performance gains over ERNIE X1 with factual accuracy up 34.8% https://t.co/6LkjZqvRS0 ERNIE X1.1 just launched Unveiled at Wave Summit 2025"
X Link 2025-09-15T12:06Z [---] followers, [--] engagements
"1/5 The Wan-Animate paper is out With just a single character image and a reference video Wan-Animate can seamlessly replace the original characterperfectly capturing the targets expressions movements and the scenes lighting and color tones. The result is AI-generated video that feels natural consistent and free of visual artifacts. So how does Wan-Animate achieve this👇👇👇 paper: #Wan #WanAnimate https://arxiv.org/abs/2509.14055v1 https://arxiv.org/abs/2509.14055v1"
X Link 2025-09-18T11:41Z [---] followers, [---] engagements
"5/5 Finally the Wan team showcased a diverse range of results produced by Wan-Animate highlighting its stability controllability and state-of-the-art performance"
X Link 2025-09-18T11:44Z [---] followers, [---] engagements
"🚀ModelScope FlowBench is now live A local-cloud collaborative platform for text / image / 3D / video / audio workflows. ✅Free cloud compute run QwenImage Wan2.2 Flux SD and more with zero GPU hassle. ✅ One-click workflow cloning copy pro pipelines instantly. Change outfits swap background figurine effects all in a single prompt. ✅ 10000+ built-in LoRAs choose switch and apply with ease. ✅ Drag-and-drop nodes + real-time preview build complex pipelines with generation editing and pose control all in one. 💻 For Free jump in today Local client download in the comments #Aiart #AITools"
X Link 2025-09-25T08:33Z [---] followers, [----] engagements
"@nhl_desktop thanks for letting me know try links below😎 🍎 💻 https://cdn-muse-cn-1.modelscope.cn/flowbench/release/win/FlowBench-0.5.0-beta-20250923-2111.exe https://cdn-muse-cn-1.modelscope.cn/flowbench/release/mac/FlowBench-0.5.0-beta-20250923-2111.dmg https://cdn-muse-cn-1.modelscope.cn/flowbench/release/win/FlowBench-0.5.0-beta-20250923-2111.exe https://cdn-muse-cn-1.modelscope.cn/flowbench/release/mac/FlowBench-0.5.0-beta-20250923-2111.dmg"
X Link 2025-09-25T08:59Z [---] followers, [--] engagements
"@nhl_desktop not yet but we'll certainly consider it"
X Link 2025-09-25T09:06Z [---] followers, [--] engagements
"🎉ModelScope DiffSynth-Studio Hits [-----] GitHub Stars Huge thanks for your incredible support 🔥 We've completely redesigned the inference and training pipelines for leading Diffusion Models like Qwen-Image Wan Hunyuan and Flux. This delivers efficient dynamic VRAM management and highly flexible model training making it easy for consumer-grade GPUs to handle large models 🌟 Key Highlights: - The ModelScope Civision has leveraged DiffSynth-Studio to generate 3000+ Qwen-Image LoRA models. - Our versatile training framework has rapidly incubated multiple Qwen-Image extension models significantly"
X Link 2025-09-28T09:50Z [---] followers, [---] engagements
"DeepSeek-V3.2-Exp is hereopen-source and blazing fast 🚀 - Tech: Powered by DeepSeek Sparse Attention delivering huge speedups on long-context training & inference with no quality tradeoff. - Performance: Matches V3.1-Terminus across top public benchmarks. - Price: Serving costs slashedAPI now over 50% cheaper effective today 🔗Get the model and try it out on ModelScope: 🔗Paper link: #DeepSeek #LLM https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2-Exp"
X Link 2025-09-29T11:06Z [---] followers, [---] engagements
"1/4 🤯 The Embodied AI "ChatGPT Moment" is 1-2 years away. But what's the path The ModelScope Embodied AI Developer Forum saw experts from DAMO Academy BAAI OpenLoong Ant Group Manycore X Square Robot Linkerbot and more unveil their roadmap for the future. The consensus: Data is the new bottleneck and open source is the solution. Top [--] breakthroughs and key takeaways in this thread 👇🧵"
X Link 2025-10-11T09:47Z [---] followers, [---] engagements
"4/4 The Community & Commercial Tipping Point: 🚀 OpenLoong: Launched China's first full-stack open-source full-size Humanoid Robot (1.85m [--] DoF) & the low-cost NanoLoong educational kit. 💸 Consumer Price Point: Experts predict a reliable multi-task robot at the [----] RMB price point could be the catalyst for mass adoption"
X Link 2025-10-11T10:43Z [---] followers, [---] engagements
"@iSaif_ai @Merjic_AI @Alibaba_Qwen Yes but make sure to choose Qwen-Image-Edit v1"
X Link 2025-10-17T13:03Z [---] followers, [---] engagements
"🔥 Shanghai AI Laboratory open-sources InternVL3.5 🚀 ✨ Highlights: [--] model sizes (1B 241B params) dense + MoE first open multimodal LLM built on GPT-OSS. 241B model tops benchmarks: MMStar [----] OCRBench [----] beats GPT-5. AIME25 [----] MMLU-Pro [----] SOTA multimodal reasoning. Cascade RL boosts avg reasoning +16 pts; 241B model hits [----] beating Claude-3.7-Sonnet. New ViR & DvD slash latency [--] (369 ms [--] ms); Flash model keeps 100% perf at 50% seq len. Upgraded agents: GUI embodied AI SVG. New SOTA on ScreenSpot (92.9) VSI-Bench (69.5) SGP-Bench (70.6). 💻 Download here: 🌐 Try it: #Multimodal"
X Link 2025-09-01T10:55Z [----] followers, [---] engagements
"🤯 [---] Token/S on a MacBook Yes you read that right Shaohong Chen just fine-tuned the Qwen3-0.6B LLM in under [--] minutes using Apple's MLX framework. This is how you turn your MacBook into a serious LLM development rig. A step-by-step guide and performance metrics inside 🧵 https://mp.weixin.qq.com/s/5sIc41fE7LkEv4ytJ7j8Eg https://mp.weixin.qq.com/s/5sIc41fE7LkEv4ytJ7j8Eg"
X Link 2025-10-13T12:01Z [---] followers, 13.6K engagements
"🚀Qwen3-VL Ultimate Guide: FREE API & Fine-Tuning Tutorial We've packed two must-have skills into one easy-to-follow guide: [--] Free Qwen3-VL API Access: Start building without spending a dime. [--] ms-swift Fine-Tuning Tutorial: Customize Qwen3-VL easily with your own data using MS-Swift. Save the long image (it's a 💰 goldmine) #qwen3VL #qwen #msswift"
X Link 2025-10-15T05:32Z [---] followers, [----] engagements
"1/3 🚀 Introducing Face-to-Photo by DiffSynth-Studio & @Merjic_AI Transform ordinary face photos into stunning high-fidelity portraits. It's now open-source 👍 Built on Qwen-Image-Edit @Alibaba_Qwen the Face-to-Photo model excels at precise facial detail restoration. Unlike previous models (e.g. InfiniteYou) it captures fine-grained facial features across angles sizes and positions producing natural aesthetically pleasing portraits. 🌟 Local deployment/usage tutorial in the thread 🧵 Model Try it: https://modelscope.cn/aigc/imageGenerationtab=advanced&imageId=17008179"
X Link 2025-10-17T06:19Z [---] followers, 21.7K engagements
"@Merjic_AI @Alibaba_Qwen 3/3 💻 Experience it in FlowBench: Use the Image-to-Image Generation node select Qwen-Image-Edit v1 + DiffSynth-Studio/Qwen-Image-Edit-F2P LoRA generate high-quality portraits with one click. https://cdn-muse-cn-1.modelscope.cn/flowbench/release/win/FlowBench-0.5.0-beta-20250929-1751.exe https://cdn-muse-cn-1.modelscope.cn/flowbench/release/mac/FlowBench-0.5.0-beta-20250929-1751.dmg https://cdn-muse-cn-1.modelscope.cn/flowbench/release/win/FlowBench-0.5.0-beta-20250929-1751.exe"
X Link 2025-10-17T06:21Z [---] followers, [---] engagements
"@crazytoolman @Merjic_AI @Alibaba_Qwen we'd love to get your feedback on this😀"
X Link 2025-10-17T07:52Z [---] followers, [---] engagements
"1/10 🤖 Deep Research Agents are redefining search Driven by OpenAI & Google the Autonomous Research Paradigm is officially mainstream. But what does a top-tier agent look like under the hood This deep dive answers: - Core definitions & capabilities of Deep Research - Frontier architecture & iterations - Engineering insights & commonalities of mainstream architecture and design We break down Architectures Core Techniques (SFT/RL) and lessons from Perplexity DeerFlow and Tongyi DeepResearch 👇🧵 Full Report (Written by LiGongsheng from Modelscope Team): #DeepResearch #AIAgent"
X Link 2025-10-21T07:39Z [---] followers, [---] engagements
"3/10 Deep Research Agent architectures fall into two categories based on LLM autonomy: Static Workflows (fixed human-defined pipelines) offer stability and are easier to implement but lack generalization; Dynamic Workflows (model-driven planning execution and reflection) provide superior flexibility and generalization for complex tasks but demand higher LLM capability and face instability challenges. A hybrid approach is often used in practice to balance stability and agility"
X Link 2025-10-21T07:41Z [---] followers, [--] engagements
"4/10 Dynamic workflows are split between Single-Agent (relying on a single LLM's long-context reasoning for planning/execution/reflection ideal for end-to-end RL e.g. Agent-R1) and Multi-Agent (allocating sub-tasks to specialized agents mimicking teamwork offering scalability e.g. deerflow). Multi-Agent systems face challenges in coordination and context management while Single-Agent designs require top-tier base models"
X Link 2025-10-21T07:41Z [---] followers, [--] engagements
"7/10 Four main techniques are employed to optimize DR Agents: Prompt Engineering (e.g. ReAct CoT) Supervised Fine-Tuning (e.g. Open-RAG AUTO-RAG) Reinforcement Learning (RL) (e.g. Agent-R1 WebThinker) and Non-parametric Continuous Learning (e.g. Case-Based Reasoning/CBR used in AgentRxiv). Specifically Tongyi DeepResearch uses a two-stage SFT+RL curriculum training to achieve state-of-the-art results on multiple benchmarks (e.g. HLE=32.9 BrowseComp-EN=43.4)"
X Link 2025-10-21T07:42Z [---] followers, [--] engagements
"8/10 Open-Source Deep Dive. Two open-source projects offering key lessons: Deep Research (17k stars): Uses a recursive search tree (custom depth/breadth) + LLM Distillation to beat context window limits. Great for cost control. DeerFlow (15.2k stars): A robust Multi-Agent system with Human-in-the-Loop (plan modification) + Global State Management for seamless context flow"
X Link 2025-10-21T07:42Z [---] followers, [--] engagements
"9/10 Key engineering insights from current DR Agent development: - Dynamic Adaptation: Continuously re-evaluate which parts of the workflow the LLM should autonomously manage as model capabilities improve; - Iterative Search: Design progressive search pipelines where new queries are generated based on previous "learnings/findings" to avoid redundancy; - Clean Context: Deliver structured "learnings/findings" instead of raw text to subsequent agents for better stability and lower cost; - Human-in-the-loop: Implement simple yet critical mechanisms for intent clarification and plan modification"
X Link 2025-10-21T07:43Z [---] followers, [--] engagements
"🤯 MIT License + 5-Minute Coherence + 10x Speed Boost. Meituan just open-sourced LongCat-Video (13.6B) a SOTA video base model that's a serious contender for the World Model race. 🎥 The Breakthrough: Natively generates up to 5-minute continuous videos by pre-training on continuation tasks solving quality decay and physics adherence. Its better than most open-source models and even challenges Google's Veo3 in core metrics. ⚡ The Efficiency: Get 10x faster inference on a single H800. This is powered by Block-Sparse Attention (BSA) which cuts computation to 10% and sampling distillation (50"
X Link 2025-10-27T10:47Z [----] followers, [---] engagements
"1/3 🚀 Open-Source #1 + 10B Activation + Agent Mastery. Introducing MiniMax-M2 the MoE model (230B total 10B activated) built specifically to master complex Coding & Agent workflows. 🏆 Ranked #1 among global open-source models by Artificial Analysis for general intelligence. 👇 Why this lightweight powerhouse is the most efficient choice for your next Agent project. 🧵"
X Link 2025-10-27T11:31Z [---] followers, [---] engagements
"3/3 End-to-End Coding & True Agent Intelligence. M2 is not just a coder; it's a developer agent. It handles multi-file projects auto-fixes code via test verification and excels on SWE-Bench and Terminal-Bench. For Agent work M2 masters complex toolchains (Shell Python Browser) and showed the crucial ability for self-correction and recovery in the BrowseComp benchmark"
X Link 2025-10-27T11:32Z [---] followers, [---] engagements
"New SOTA GUI Grounding model UI-Ins is here from @Ali_TongyiLab & RUC🚀It's the first to treat user commands as dynamic reasoning pathways enabling human-like multi-perspective intent understanding. 📊 Achieves a 74.1% Task Success Rate in AndroidWorld surpassing Gemini [---] Computer Use (69.7%) 🌍 Open Source (7B/32B models) with full SFT/RL code. Dive in👇🧵 🤖 UI-Ins-7B: UI-Ins-32B: 📄 arXiv: https://modelscope.cn/papers/2510.20286 https://modelscope.cn/models/Tongyi-MiA/UI-Ins-32B https://modelscope.cn/models/Tongyi-MiA/UI-Ins-7B https://modelscope.cn/papers/2510.20286"
X Link 2025-10-30T13:42Z [---] followers, 25.6K engagements
"4/4 Size Isn't Everything: The 7B Emergent Win On MMBench-GUI L2s Advanced subset (tasks requiring implicit intent) the UI-Ins-7B model beats Qwen2.5-VL-7B by 159% Why the huge jump Structured reasoning generic free-form "lets think step by step" CoT which actually hurt grounding performance. The UI-Ins 4-view reasoning framework proves how you reason matters more than how large your parameter count is. 🧠"
X Link 2025-10-30T13:43Z [---] followers, [---] engagements
"🚀 Introducing LongCat-Flash-Omni a 560B-parameter (27B activated) open-source omni-modal MoE model excelling at real-time audio-visual interaction. Built on LongCat-Flashs high-performance shortcut-connected MoE architecture with zero-computation experts plus efficient multimodal perception & speech reconstruction modules. ✅ Unified offline multimodal understanding + real-time AV interaction ✅ 128K-token context window ✅ Trained via curriculum-inspired progressive strategy strong in both uni- and multi-modal tasks ✅ MIT-licensed full weights open-sourced 📊 SOTA results across OmniBench"
X Link 2025-11-01T09:05Z [----] followers, 12.6K engagements
"🚨@Kimi_Moonshot just open-sourced Kimi Linearthe FIRST hybrid attention architecture that BEATS full attention under fair comparisons. No training tricks. No parameter inflation. Just better engineering. Why this changes everything for LLM inference 👇 ✨ CORE INNOVATION: Kimi Delta Attention (KDA) Channel-wise gating (not head-wise) for precise memory management Specialized DPLR transition matrices for hardware efficiency First production-ready linear attention that doesn't sacrifice quality 100% PyTorch no custom CUDA required ⚡ REAL-WORLD GAINS (48B param model): 128K context: [----] faster"
X Link 2025-11-03T05:50Z [---] followers, [----] engagements
"🎨 Big news from ModelScope 🎨 One of our community developers just trained the Qwen-Image-Edit-2509 LoRA using ModelScopes no-code training tools and its now trending on Hugging Face 🚀 Were thrilled to see creatives turning their ideas into powerful LoRA models with Qwen-Image-Edit. Even better is now open globally 🌍Free training for Qwen-Image and other T2I models is live with Qwen-Image-Edit LoRA training support coming later this month. 💫Try it now: https://modelscope.ai/civision/modelTraining http://ModelScope.ai https://modelscope.ai/civision/modelTraining http://ModelScope.ai"
X Link 2025-11-04T06:32Z [---] followers, [----] engagements
"Thanks @CGTNOfficial for showcasing the incredible open-source AI revolution From mind-reading glasses to knowledge-hungry robots the future is already here. ModelScope is one of the open communities where this revolution begins with every developer who dives in. Watch the video and join the movement When Code Has No Secrets Who Defines the Future #PulseofModernization #APEC2025 https://t.co/te0QPqS0W6 When Code Has No Secrets Who Defines the Future #PulseofModernization #APEC2025 https://t.co/te0QPqS0W6"
X Link 2025-11-05T08:24Z [---] followers, [---] engagements
"1/4Just fine-tuned Qwen3-8B on Huawei Ascend NPU for quant finance coding tasks 🔥 🤖Inspired by RD-Agentinstead of calling cloud LLMs to convert quant formulas code Yuan Guo fine-tuned Qwen3-8B to do it locally: faster cheaper and purpose-built for finance. Built a specialized model that converts complex LaTeX quant factor formulas directly into production-ready Python functions. Achieved +20 pts over base model on code correctness robustness & style.🚀 Key stack: Huawei Ascend 910B3 (64GB) LLaMA Factory @llamafactory_ai + LoRA SwanLab for experiment tracking MindNLP for deployment Best"
X Link 2025-11-06T04:45Z [---] followers, [---] engagements
"Special Release Who Codes the Future New data reveals the staggering scale of Chinas AI ecosystem: 9.4M+ developers 30M+ open-source projects 346K+ AI companies in H1 [----]. 🚀This isnt just growthits a full-stack innovation machine shaping global trends. D-3 until The Next Wave: Who Codes the Future The AI Developer Ecosystem Report by @ModelScope2022 x @ZhihuFrontier Every line of code carries a worldview. Every developer a choice about the future. #AI #Developers #ModelScope #Zhihu #ChinaAI"
X Link 2025-11-07T02:39Z [----] followers, [---] engagements
"🔥 BREAKING: @StepFun_ai just open-sourced Step-Audio-EditXthe world's first LLM-grade audio editing model 🎧✨ Control emotion speaking style AND paralinguistic elements (breaths laughs sighs) through simple text prompts. Zero-shot TTS with multilingual support (EN/CN + dialects) in a unified 3B-parameter architecture. ✅ Single GPU deployable ✅ Apache [---] license ✅ Iterative editing capabilities No more patchwork pipelinesaudio creation just got an LLM upgrade. Model: GitHub: #AI #AudioAI #OpenSource #LLM #GenerativeAI #SpeechTech #StepFunAI http://github.com/stepfun-ai/Step-Audio-EditX"
X Link 2025-11-08T07:11Z [---] followers, [----] engagements
"🚀 Meet @Meituan_LongCat 's UNO-Benchthe first unified benchmark that finally cracks the black box of omni-modal AI. The breakthrough: UNO-Bench reveals a compositional law: 🔹 Weak models hit a bottleneck 🔹 Strong models get superlinear synergy (P_omni (P_audio P_visual)2.19) ✅ [----] human-curated omni samples (98% cross-modal solvable) ✅ Novel multi-step open-ended QA + 95% accurate auto-grading ✅ 90% faster eval with 98% consistency vs. [--] public benchmarks If youre building or evaluating multimodal modelsthis is your new gold standard. 🔗 #AI #Multimodal #Benchmark #LLM #OmniModel"
X Link 2025-11-09T07:51Z [---] followers, [----] engagements
"🚀 ModelScope just launched Mcore-Bridge making high-performance Megatron training as easy as Transformers. Train MoE models up to 10x faster than DeepSpeed with seamless integration into vLLM SGLang & Hugging Face ecosystems. No more complex setup no manual weight conversion just plug & play with native safetensors support. 🔥 Key highlights: ✅ 10x speedup on MoE models (e.g. Qwen3-MoE) ✅ Native safetensors I/O no format hassles ✅ Full LoRA support + bidirectional PEFT compatibility ✅ Smooth Megatron vLLM weight sync for RLHF (GRPO/GKD) ✅ Python API & CLI for flexible debuggable workflows ✅"
X Link 2025-11-10T09:44Z [---] followers, [----] engagements
"🔥 3B active params. Full multimodal thinking. ERNIE-4.5-VL-28B-A3B-Thinking is here a lightweight vision-language model with deep reasoning image zoom search grounding & video understanding. ✅ Apache [---] Commercial use allowed ✅ ModelScope vLLM FastDeploy ready ✅ SFT/LoRA support via ERNIEKit Outperforms expectations across visual reasoning STEM and real-world tasks all at minimal cost. 👉 Try it now: #AI #MultimodalAI #LLM #OpenSource #AIAgent #ComputerVision #PaddlePaddle https://modelscope.cn/models/PaddlePaddle/ERNIE-4.5-VL-28B-A3B-Thinking"
X Link 2025-11-11T10:44Z [---] followers, [----] engagements
"🧠 Turn expertise into plug-and-play AI skills. 🚀MS-Agent now implements Anthropics Agent Skills protocol a modular way to give agents domain knowledge via structured folders of instructions scripts and resources. ❗Why it matters As agents grow more capable we need composable portable and scalable ways to inject specialized knowledge beyond simple tool calling. Skills solve this by packaging human workflows as reusable onboarding guides for AI. 🔧 How it works - Skills live in folders with (YAML + Markdown) optional scripts/ resources/ - MS-Agent loads them on-demand using 4-level context"
X Link 2025-11-11T12:59Z [---] followers, [---] engagements
"Large models have made development easier and sped up progress but theyve also birthed a new role: the LLM Toilet Paper Engineer.🤣 Were now living in the 80-Point Crisis where everything almost works but no one dares to ship it. Executives see the magic. They dont see the mess behind it. Jinhui a contributor from @ModelScope2022 and @ZhihuFrontier shared his insights during a rencent panel -- and what do you think"
X Link 2025-11-12T05:39Z [---] followers, [----] engagements
"KookYan just released Kook_Qwen_zshx_v2 a powerful LoRA built on Qwen-Image🚀 👇Example prompts are in the comments below Its designed to SMASH the boundary between real-life photorealism and CG fantasy. 🧙This powerful LoRA masters the "Real-life cinematic quality + pure fantasy scene" fusion serving as a massive shortcut for film concept artists game developers and virtual content creators looking to produce "cinematic real-fantasy fusion works" efficiently 💡 Key Features for Creators: - Realism in Fantasy: Achieve the "real person texture + fantasy scene" fusion without the typical"
X Link 2025-11-13T12:30Z [---] followers, [----] engagements
"8. Asian female real person. "Mandatory specified art style": "Realistic style photo taken with a SLR camera" "Lens": "Medium shot equivalent 85mm focal length shallow depth of field eye-level perspective slightly front-left angle highlighting the subject's silhouette and details" "Composition": "Subject slightly right of center vertical composition foreground is the upper body of the character background is a rough rock wall; character occupies about two-thirds of the image height cropped below the shoulders emphasizing the torso and costume details; background blurred but retaining texture"
X Link 2025-11-13T12:34Z [---] followers, [---] engagements
"9. A sexy girl dancing movements are large and beautiful wearing a yellow-green and gold colored Hanfu off-the-shoulder exquisite hair accessories full body beautiful leg lines low-cut sexy Da Vinci color grading transparency real person realistic background is a dark partially light-transmitting golden palace slight low angle slightly tilted lens delicate fair skin with a dewy luster cold and arrogant eyes natural messy strands of hair falling down subtle reflection on hair strands black smooth long hair proud figure full and upright breasts golden characters "Chang Le Wei Yang" (Eternal Joy"
X Link 2025-11-13T12:35Z [---] followers, [---] engagements
"11. surreal realistic_photo photography high_contrast hyper_detailed low_saturation "caption": "A red-haired woman wearing Hanfu a large-sleeved Hanfu of the Wei and Jin style holding a long scepter the scepter is taller than her the head of the scepter is a copper ring strings of small bells hang from the copper ring swaying in the wind as if hearing the crisp sound of the bells. She faces the camera tilting her head smiling slightly. The skirt is long and extends to the ground. Emphasizing the messy beauty of dynamic hair strands extremely high visual appeal Rembrandt light reflection"
X Link 2025-11-13T12:36Z [---] followers, [---] engagements
"🔥 Big shoutout to @Eigen_AI_Labs for this release Our DiffSynth-Studio team saw it and theyre very happy. 😄 Meet Eigen-Banana-Qwen-Image-Edit: ⚡ 4-step pro-quality edits (down from 50+) 🖋 Pixel-perfect text + semantic control 🧩 Open LoRA weights runs natively in DiffSynth-Studio & Diffusers 🛠 Built on EigenPlatform + Pico-Banana-400K eigen-banana-qwen-image-edit Eigen-Banana-Qwen-Image-Edit is a LoRA (Low-Rank Adaptation) checkpoint for the Qwen-Image-Edit model optimized for fast high-quality image editing with text prompts. This model enables efficient text-guided image transformations"
X Link 2025-11-14T09:28Z [---] followers, [----] engagements
"🚀Meet MemOS: the first Memory Operating System for LLMs. Open source. Built for agents with long-term memory. ✅ Persistent searchable evolving memories ✅ Standard MCP I/O no SDKs no plugins ✅ Works with ModelScope 14.9k+ calls in [--] week 🔗Try it live: 🔗Star on GitHub: Let memory become AIs new compute layer. #AI #LLM #MemOS #MCP http://github.com/MemTensor/MemOS https://modelscope.cn/mcp/servers/MemTensor/MemoryOperatingSystem http://github.com/MemTensor/MemOS https://modelscope.cn/mcp/servers/MemTensor/MemoryOperatingSystem"
X Link 2025-11-17T07:44Z [---] followers, [----] engagements
"1/4 🚀 Meet AgentEvolver from @Ali_TongyiLab : AI agents that learn like humans no hand-labeled data needed. A 7B model beats a 14B baseline. 57.6% success rate vs. 29.8% with 50% fewer parameters. The team have open-sourced everything: 🔗 Code: 📄 Paper: No more costly RL loops or random exploration. ✅ Self-questioning generates its own tasks ✅ Self-navigating reuses experience smarter paths ✅ Self-attributing rewards smart actions not just luck Built on LLMs optimized for efficiency. Faster adaptation. Less data. More autonomy. "The future of AI agents isnt about bigger models. Its about"
X Link 2025-11-18T09:05Z [---] followers, [----] engagements
"3/4 1Self-Questioning Instead of waiting for humans to define tasks AgentEvolvers LLM looks at its environment and asks: What can I try next It autonomously generates novel meaningful challenges no manual dataset needed. Think of it as curiosity encoded. 2Self-Navigating Every failure every success its not wasted. The agent summarizes past experiences into reusable mental maps then blends them with fresh exploration. No more repeating the same mistakes. It learns from history not just trial-and-error. 3Self-Attributing In long tasks you rarely know which step led to success. AgentEvolver runs"
X Link 2025-11-18T09:07Z [---] followers, [---] engagements
"🔥The LoRAs built on Qwen-Image-Edit-2509 are blowing upand have dominated Hugging Faces Trending list. All were trained on ModelScope and are available via free API: 📷 Multiple-angles rotate product shots like a [---] camera 💡 Light_restoration erase shadows enhance lighting naturally 🛒 Fusion & White_to_Scene seamlessly place products into real-world scenes (no Photoshop needed) 🕯 Relight simulate soft window glow studio lighting and more 👥 Multiple-characters generate entire scenes with consistent characters from a single image 🔧 Workflow expand retouch virtual try-on pose control all"
X Link 2025-11-19T12:08Z [----] followers, [----] engagements
"🚀 @TencentHunyuan Just dropped: HunyuanVideo [---] the lightweight SOTA video gen model that runs on a 14GB consumer GPU. 8.3B params. DiT architecture. Realistic 510s 480p/720p videos. Upscale to 1080p. ✅ Mid-prompt control (English/Chinese) ✅ Image-to-video with perfect consistency ✅ Cinematic camera moves text in video physics-aware motion ✅ SSTA attention = 3x faster inference than 20B+ models No A100 needed. Just deploy & play. ModelScope: GitHub: https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5 https://www.modelscope.cn/models/Tencent-Hunyuan/HunyuanVideo-1.5"
X Link 2025-11-21T05:08Z [----] followers, [----] engagements
"Meet Nex by NEX-AGI a non-thinking model built for agents that crushes it in coding tool use and roleplay 🚀 ✅ SOTA among open models on Tau2-Bench BFCL V4 GAIA2 ✅ Top-tier in frontend vibe coding and mini-program/backend dev (human eval confirmed) ✅ Plug-and-play with Claude Code Cursor etc. 👉 Expolore Nex: 🔥 Free for now on SiliconFlow: https://www.siliconflow.com/models/deepseek-v3-1-nex-n1 https://modelscope.cn/organization/nex-agi https://www.siliconflow.com/models/deepseek-v3-1-nex-n1 https://modelscope.cn/organization/nex-agi"
X Link 2025-11-24T06:04Z [----] followers, [----] engagements
"🔥 @TencentHunyuan just open-sourced HunyuanOCR a 1B-parameter end-to-end multimodal OCR model that beats Gemini Qwen-VL and commercial APIs across 10+ benchmarks. ✅ SOTA on OmniDocBench (94.1) ✅ Full document parsing: LaTeX formulas HTML tables reading-order Markdown ✅ 100+ languages 14EN/ZH photo translation (ICDAR2025 small-model champ) ✅ JSON field extraction (receipts IDs) bilingual subtitle extraction ✅ Single inference no cascades lightweight & production-ready All with just 1B params 🚀 Try it now: 🔗 🔗 https://github.com/Tencent-Hunyuan/HunyuanOCR"
X Link 2025-11-25T07:42Z [----] followers, [----] engagements
"🚀Bytedance-Research just dropped: VeAgentBench a new benchmark dataset to evaluate real-world agentic capabilities ✅ [---] open Qs (484 total) across legal finance edu & personal assistant ✅ Tests tool use RAG memory & multi-step reasoning ✅ Comes with runnable agents built on veADK (Volcengines full-stack agent framework) Perfect for devs pushing agentic AI beyond toy demos. 🔗 Dataset: 📜 CC BY-NC [---] (non-commercial) https://modelscope.cn/datasets/bytedance-research/veAgentBench/summary https://modelscope.cn/datasets/bytedance-research/veAgentBench/summary"
X Link 2025-11-25T13:21Z [----] followers, [----] engagements
"Introducing Z-Image: a powerful highly efficient 6B-parameter image generation model with three specialized variants 🎨 🚀 Z-Image-Turbo Distilled for speed & quality Just [--] NFEs matches or beats top competitors ⚡Sub-second latency on H800 GPUs Runs smoothly on 16GB VRAM consumer cards Photorealism + bilingual text (EN/ZH) + strong instruction following Try it now: 🔗 Demo: 🔗 Model: https://modelscope.cn/models/Tongyi-MAI/Z-Image-Turbo https://modelscope.cn/aigc/imageGeneration https://modelscope.cn/models/Tongyi-MAI/Z-Image-Turbo https://modelscope.cn/aigc/imageGeneration"
X Link 2025-11-27T01:55Z [----] followers, 21.5K engagements
"Free Z-Image API is live 🎉 The open-source Z-Image Turbo model is now 100% free to call via API on ModelScope🔥 AND Free compute included ✨ How to use it: [--] Try it instantly: Go to Z-Image is the default model - Quick Mode: Fast results zero config - Advanced Mode: Full control for fine-tuned outputs [--] Integrate it anywhere: Visit 👉 copy the Python code and run it: - Locally or in Jupyter - In your own apps - With ComfyUI or other AI workflows Start generating now #ZImage_Turb https://modelscope.cn/models/Tongyi-MAI/Z-Image-Turbo https://modelscope.cn/aigc/imageGeneration"
X Link 2025-11-28T07:02Z [----] followers, 50.2K engagements
"🚀 Meet GELab-Zero-4B-preview a 4B multimodal GUI agent from @StepFun_ai fine-tuned for Android & runs locally on consumer hardware ✅ Zero-shot UI automation: taps swipes typing all from screenshots/video ✅ Handles complex cross-app tasks (food rides shopping social) ✅ No app-specific training needed works on unseen dynamic interfaces ✅ Plug-and-play infra: ADB deps recording/replay included Built on Qwen3-VL-4B-Instruct. Small model. Big autonomy. Local. Visual. General. 🔥 🤖Model: https://modelscope.cn/models/stepfun-ai/GELab-Zero-4B-preview"
X Link 2025-12-01T02:56Z [----] followers, [----] engagements
"DeepSeek-V3.2 is live 🚀 @deepseek_ai and its the most capable open-agent LLM yet. ✅ Matches GPT-5 on reasoning ✅ First open model to combine chain-of-thought + tool calling ✅ Trained on 85K+ synthetic agent tasks SOTA on SWE-Bench Tool Decathlon & more ✅ Built with DSA: [--] cheaper inference at 128K And the research variant V3.2-Speciale just won gold medals in IMO [----] IOI [----] ICPC WF on par with Gemini-3.0-Pro.🏆 📄 Tech Report: 🔗 Open weights on ModelScope: https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2 https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.2-Speciale"
X Link 2025-12-01T11:43Z [----] followers, 14K engagements
"🔥 The Qwen-Image-Edit-2509 API is now live and 100% free to use ✨ Free inferences refresh daily 🎁 New users get free GPU hours on signup 🚀 Ready to build Try it on ModelScope: https://modelscope.cn/models/Qwen/Qwen-Image-Edit-2509 https://modelscope.cn/models/Qwen/Qwen-Image-Edit-2509 https://modelscope.cn/models/Qwen/Qwen-Image-Edit-2509 https://modelscope.cn/models/Qwen/Qwen-Image-Edit-2509"
X Link 2025-12-02T08:10Z [----] followers, 31K engagements
"🚀 Ovis-Image (7B) is live on ModelScope ✅Delivers frontier-level text renderingon par with 20B-class models like Qwen-Image and even competitive with GPT-4o on text-heavy tasks. ✅Sharp layout-aware output for posters banners logos UI mocks and infographics. ✅Runs fast and leandeployable on a single high-end GPU. Small model. Big text fidelity. 👉 https://modelscope.cn/models/AIDC-AI/Ovis-Image-7B https://modelscope.cn/models/AIDC-AI/Ovis-Image-7B"
X Link 2025-12-03T02:14Z [----] followers, [----] engagements
"🚀 Introducing Flowra by ModelScope & WULI The open-source engine behind FlowBench makes building AI workflows as easy as snapping LEGO bricks. ✅ Unified handling of images audio video & 3D ✅ DAG-based execution w/ smart caching & distributed scaling ✅ One-line ModelScope model integration ✅ Full dev toolkit: flowra create build debug deploy Turn your ML model into a visual node in minutes. ✨ No more dependency hell. No more blind debugging. 🔗 GitHub: 📥 FlowBench client: https://modelscope.cn/flowbench/download https://github.com/modelscope/flowra https://modelscope.cn/flowbench/download"
X Link 2025-12-04T11:01Z [----] followers, [----] engagements
"🔥 New I2V breakthrough: SteadyDancer generates ultra-coherent human animation videos while perfectly preserving the first frames identity. ✅ Solves identity drift & motion artifacts ✅ Uses smart conditional harmonization + pose-aware modules ✅ Trains faster + outperforms SOTA on FVD VBench SSIM No more uncanny valleyjust smooth controllable reference-accurate motion. 🤖 Model: 📄 Paper: https://modelscope.cn/papers/2511.19320 https://modelscope.cn/models/MCG-NJU/MCG-NJU-SteadyDancer-14B https://modelscope.cn/papers/2511.19320 https://modelscope.cn/models/MCG-NJU/MCG-NJU-SteadyDancer-14B"
X Link 2025-12-05T02:57Z [----] followers, [----] engagements
"🚀Introducing LongCat-Image from @Meituan_LongCat a groundbreaking open-source bilingual (Chinese-English) image generation foundation model 🐾 ✨ Just 6B params yet outperforms much larger models in efficiency & quality ✨ Best-in-class Chinese text rendering accurate stable and covers the vast majority of the Chinese lexile ✨ Striking photorealism powered by a novel data & training strategy Built for developers optimized for real-world use. Open accessible and multilingual from the ground up. 🤖Model: https://modelscope.cn/models/meituan-longcat/LongCat-Image"
X Link 2025-12-06T05:05Z [----] followers, [----] engagements
"🚀 NewBieAI-Lab drops NewBie-image-Exp0.1 a 3.5B open-source ACG-native DiT model built for precise fast and high-quality anime generation. ✅ 3.5B params (8GB VRAM friendly RTX [----] ✅) ✅ Dual text encoders: Gemma-3-4B-it + Jina CLIP v2 deep prompt understanding ✅ XML-structured prompts for per-character control (no more outfit swapping) ✅ FLUX.1-dev 16-ch VAE buttery skin fabric metal ✅ 20-step inference LoRA-friendly Apache-2.0 + non-commerical license ✅ Trained on 10M+ anime images w/ XML annotations rock-solid multi-character scenes 🤖Models: ⚡ 40% faster than 8B+ models yet handles"
X Link 2025-12-07T05:47Z [----] followers, 43.2K engagements
"🚀 Introducing Live Avatar: real-time streaming infinite-length interactive avatar video generation Powered by a 14B-parameter diffusion model [--] FPS on 5H800 GPUs (4-step sampling). Chat naturally via mic & camyour avatar responds instantly with lifelike video. Paired with Qwen3-Omni it enables fully autonomous face-to-face AI conversations. Watch two agents talkstreaming in real time for 10000+ seconds. 👀 Useful Resources👇 YouTube: Model: Blog: Paper: Code: https://github.com/Alibaba-Quark/LiveAvatar https://modelscope.cn/papers/2512.04677 https://liveavatar.github.io/"
X Link 2025-12-08T03:34Z [----] followers, [----] engagements
"Excellent work from @dx8152 🔗Qwen-Edit-2509-Light-igration: https://modelscope.ai/models/dx8152/Qwen-Edit-2509-Light-igration Love seeing what the community builds with @ModelScope2022 . @dx8152 just dropped a game-changing Light Migration LoRA for Qwen-Image-Edit-2509. It solves the "secondary lighting" headache perfectly. Incredible work. 👏 https://t.co/5IYTfR53uD https://modelscope.ai/models/dx8152/Qwen-Edit-2509-Light-igration Love seeing what the community builds with @ModelScope2022 . @dx8152 just dropped a game-changing Light Migration LoRA for Qwen-Image-Edit-2509. It solves the"
X Link 2025-12-08T07:06Z [----] followers, [----] engagements
"🚀This 80B LLM runs on 64GB RAMthanks to GGUF + smart quantization👏 Qwen3-Next-80B-A3B-InstructGGUFModelScope 64GBPCLLMhttps://t.co/me02jHdcrY Qwen3-Next-80B-A3B-InstructGGUFModelScope 64GBPCLLMhttps://t.co/me02jHdcrY"
X Link 2025-12-08T08:02Z [----] followers, [----] engagements
"🔥BREAKING DiffSynth-Studio just dropped: Qwen-Image-i2L the first open-source suite that turns a single image into a custom LoRA 🖼➡🧠 Choose your flavor: 🎨 Style pure aesthetic extraction (2.4B) 🧩 Coarse captures content + style (7.9B) ✨ Fine 1024x1024 detail booster (7.6B pairs with Coarse) ⚖ Bias aligns output with Qwen-Images native vibe (30M) Built on SigLIP2 + DINOv3 + Qwen-VL. Ready to train LoRAs from one image Its open. Its out. Go break something cool. 💥 🤖ModelScope: 💻Code:"
X Link 2025-12-09T15:06Z [----] followers, 130.8K engagements
"🚀 Good news for Z-Image Turbo users Weve open-sourced DistillPatch LoRA to restore true 8-step Turbo speed for any LoRA fine-tuned on Z-Image Turbo. 💡 The problem: Standard LoRA training breaks Z-Image Turbos accelerationimages turn blurry at steps=8 cfg=1 even though they look fine at steps=30. ✅ We tested [--] approaches: [--] Standard SFT high quality but slow [--] Diff-LoRA fast but inflexible [--] SFT + distillation great results heavy lift [--] SFT + DistillPatch (our recommendation) plug in one LoRA instantly get 8-step Turbo performance back 🚀 Perfect if you: Already have a Z-Image Turbo LoRA"
X Link 2025-12-15T02:37Z [----] followers, 38.1K engagements
"🚀 Tongyi FUN just leveled up This release brings major upgrades across TTS & ASR: 🧠 Fun-CosyVoice3 (0.5B) Open-sourced - Zero-shot voice cloning - Local deployment & secondary development ready Try Demo: Model(Chinese): Model(Multi-Language): ⚡ Fun-ASR-Nano (0.8B) Open-sourced - Lightweight Fun-ASR variant - Lower inference cost - Local deployment & custom fine-tuning supported Model: Open fast and production-ready. ✨ https://modelscope.cn/models/FunAudioLLM/Fun-CosyVoice3-0.5B-2512 https://modelscope.cn/models/FunAudioLLM/Fun-ASR-MLT-Nano-2512"
X Link 2025-12-15T09:13Z [----] followers, 23.8K engagements
"New Z-Image-Turbo control model released 🎉 More control blocks. Inpaint mode. Better details. - 6-block ControlNet - Trained from scratch on 1M images @1328 res - Canny / HED / Depth / Pose / MLSD - Best control_context_scale: 0.65-0.80 Model: Demo: Code: https://github.com/aigc-apps/VideoX-Fun https://modelscope.cn/studios/AI-ModelScope/ZIT-Controlnet https://modelscope.cn/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union https://modelscope.cn/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union https://github.com/aigc-apps/VideoX-Fun https://modelscope.cn/studios/AI-ModelScope/ZIT-Controlnet"
X Link 2025-12-16T03:15Z [----] followers, 33.1K engagements
"🚀Model & Demo are now live on our international site Model: - - - Demo: https://modelscope.ai/studios/FunAudioLLM/Fun-CosyVoice3-0.5B https://modelscope.ai/models/FunAudioLLM/Fun-CosyVoice3-0.5B-2512 https://modelscope.ai/models/FunAudioLLM/Fun-ASR-Nano-2512 https://modelscope.ai/models/FunAudioLLM/Fun-ASR-MLT-Nano-2512 https://modelscope.ai/studios/FunAudioLLM/Fun-CosyVoice3-0.5B https://modelscope.ai/models/FunAudioLLM/Fun-CosyVoice3-0.5B-2512 https://modelscope.ai/models/FunAudioLLM/Fun-ASR-Nano-2512 https://modelscope.ai/models/FunAudioLLM/Fun-ASR-MLT-Nano-2512"
X Link 2025-12-16T09:43Z [----] followers, [---] engagements
"🤖Meet Nanbeige4-3B from Boss Zhipina 3B-parameter LLM that outperforms Qwen3-32B on math (AIME) science (GPQA) and tool calling (BFCL-V4) while matching Qwen3-30B-A3B on human preference alignment (Arena-Hard-V2). How ✅ 23T tokens of ultra-curated data ✅ Fine-grained WSD scheduler ✅ 30M+ high-quality SFT instructions ✅ Multi-stage RL + innovative distillation (DPD) ✅ Chain-of-thought reconstruction & deliberative generation It even ranks top [--] on WritingBench & EQ-Bench3beating models 100x larger like GLM-4.5 and Deepseek-R1. All models + tech report now open-source: 🔗 Weights: 📄 Paper:"
X Link 2025-12-16T11:28Z [----] followers, 25.8K engagements
"🎉 MiMo-V2-Flash FREE API is now live on ModelScope The first major release since Fuli Luo joined Xiaomiand its built for real-world agentic AI. ⚡ MiMo-V2-Flash: an open high-performance MoE model with 309B total / 15B active parameters 256K context window 150+ tokens/s generation thanks to native Multi-Token Prediction 🔥 Key wins for developers: ✅ Hybrid attention (5:1 SWA + Global) [--] less KV cache full long-context recall ✅ 73.4% on SWE-Bench Verified new SOTA for open-source models ✅ Matches DeepSeek-V3.2 on reasoning but much faster in practice ✨ API-readyperfect for building smart"
X Link 2025-12-18T02:29Z [----] followers, 16K engagements
"🚀Qwen-Image-Layered is now live on ModelScopean incredible model that can intelligently decompose any image into multiple RGBA layers 🤖 Model: 🌟 Try Demo 📑 Paper: https://modelscope.cn/papers/2512.15603 https://modelscope.cn/studios/Qwen/Qwen-Image-Layered https://modelscope.ai/models/Qwen/Qwen-Image-Layered https://modelscope.cn/papers/2512.15603 https://modelscope.cn/studios/Qwen/Qwen-Image-Layered https://modelscope.ai/models/Qwen/Qwen-Image-Layered 🎨 Qwen-Image-Layered is LIVE native image decomposition fully open-sourced ✨ Why it stands out ✅ Photoshop-grade layering Physically"
X Link 2025-12-19T15:29Z [----] followers, [----] engagements
"Its still Sunday and Im already hyped for next week. 😊The mysterious Eastern power is dropping a trilogy of magic: 🔮 Next week: A secret image-generation model 🤖 Next week: A secret code-agent model 🎙 The week after: A secret voice model Buckle upthings are about to get interesting"
X Link 2025-12-21T09:54Z [----] followers, 10.1K engagements
"🚀 Meet GLM-4.7 your new coding partner supercharged Try it now on ModelScope free API access for a limited time ✅ +5.8% on SWE-bench (73.8%) ✅ +12.9% on Multilingual coding (66.7%) ✅ +16.5% on Terminal Bench [---] (41%) ✅ Smarter agent reasoning & tool use ✅ Better UI cleaner webpages sharper slides ✅ +12.4% on complex reasoning (HLE: 42.8%)From coding to creativity GLM-4.7 delivers across the board. 🤖 Model: https://modelscope.cn/models/ZhipuAI/GLM-4.7 https://modelscope.cn/models/ZhipuAI/GLM-4.7"
X Link 2025-12-23T02:23Z [----] followers, [----] engagements
"🚀 New on ModelScope: QwenLong-L1.5 is now fully open-source A 30B model (3B active params) that matches GPT-5 & Gemini-2.5-Pro in long-context reasoning. 🔥 Key wins: ✅ +31.7 pts on OpenAIs MRCR (128K context SOTA across all models) ✅ Matches Gemini-2.5-Pro on [--] major long-QA benchmarks ✅ +9.69 on CorpusQA +6.16 on LongBench-V2 How Three breakthroughs: [--] Synthetic data at scale: 14.1K long-reasoning samples from 9.2B tokens no human labeling. Avg. length: 34K tokens (max: 119K). [--] Stable RL training: Task-balanced sampling + Adaptive Entropy-Controlled Policy Optimization (AEPO) for reliable"
X Link 2025-12-23T07:41Z [----] followers, 44.2K engagements
"@Sean60133791259 @bdsqlsz Could I ask which country you are in Well test the speed. You can DM me"
X Link 2025-12-24T02:04Z [----] followers, [--] engagements
"Merry Christmas everyone 🎄✨ 🚀Introducing BEYOND REALITY Z IMAGE 1.0: a community-tuned checkpoint merging LoRA enhancements into Z-Image optimized for film-style portrait aesthetics and high-frequency texture detail think skin pores fabric weaves and wall surfaces rendered with striking realism. ✅ Keeps Zs photorealism intact ✅ Adds cinematic color depth & tonal nuance ✅ Significantly upgrades portrait + environmental textures ✅ Runs on 8GB VRAM (FP8) tested on laptops Note: Training focused on high-quality synthetic portrait photography so complex poses or non-human textures may vary."
X Link 2025-12-24T08:00Z [----] followers, 13.5K engagements
"We shouldve shouted this louder: Youre officially missing out if you aren't using Civision 🤫 🎨 FREE Image & Video Gen & FREE LoRA Training ✨ A UI so clean you dont need a PhD to master it. Weve got the heavy hitters ready: Z-Image Qwen-Image Qwen-Image-Edit Wan2.2 and more. All the power zero cost. Right here on ModelScope: 🚀✨ https://modelscope.ai/civision/imageGeneration https://modelscope.ai/civision/imageGeneration"
X Link 2025-12-25T08:43Z [----] followers, [----] engagements
"🚀New on ModelScope: ChenkinNoob-XL v0.2 is live Independently developed by the ChenkinNoob team as a fine-tuned extension of noobai-XL-1.1 v0.2 incorporates Danbooru data up to Nov [--] 2025enhancing performance while preserving the original aesthetic. ✨ Key improvements over v0.1: ✅ Stronger character consistency ✅ Refined details (hands accessories layered outfits) ✅ Higher prompt adherence 🎨 Faithful to the noob visual DNA: clean composition solid anatomy and harmonious color palettes. ⚙ Recommended config: CFG: [--] Steps: [----] Sampler: Euler a Resolution: [----] (e.g. [-------] 10241024)"
X Link 2025-12-26T02:26Z [----] followers, 30.7K engagements
"🚀 New on ModelScope: MiniMax M2.1 is open-source ✅ SOTA in 8+ languages (Rust Go Java C++ TS Kotlin Obj-C JS) ✅ Full-stack Web & mobile dev: Android/iOS 3D visuals vibe coding that actually ships ✅ Smarter faster 30% fewer tokens with lightning mode (M2.1-lightning) for high-TPS workflows ✅ Top-tier on SWE-bench VIBE and custom coding/review benchmarks ✅ Works flawlessly in Cursor Cline Droid BlackBox and more Its not just better code its AI-native development end to end. 🔗 Model: https://modelscope.cn/models/MiniMax/MiniMax-M2.1 https://modelscope.cn/models/MiniMax/MiniMax-M2.1"
X Link 2025-12-26T08:03Z [----] followers, 17.3K engagements
"@nnnnmailcom Hi there SDXL models should be supported. Could you let me know exactly what issue you're running into"
X Link 2025-12-29T03:53Z [----] followers, [--] engagements
"🚀400+ Free APIs are now available on We currently support: ✅ DeepSeek series ✅ Qwen3 series & Qwen-Image ✅ Z-Image-Turbo and many more Check it out: Go to - Models - Filter by API-Inference. (Currently available on PC only.) 👉 Give them a spin and let us know what you think ✉DM us for support. https://modelscope.ai/modelsfilter=inference_type&page=1&tabKey=task http://modelscope.ai http://ModelScope.ai http://modelscope.ai http://ModelScope.ai https://modelscope.ai/modelsfilter=inference_type&page=1&tabKey=task http://modelscope.ai http://ModelScope.ai http://modelscope.ai"
X Link 2025-12-29T08:01Z [----] followers, 17.8K engagements
"@blankbraindead oh I see. May I know which region you are in"
X Link 2025-12-29T08:33Z [----] followers, [---] engagements
"🚀 Tencent HunYuan open-sourced HY-MT1.5 A lightweight powerhouse for translation: ✨ 1.8B model: 0.18s latency runs offline on mobile & outperforms most commercial APIs. Even rivals Gemini-3.0-Pro at 90% quality ⚡ 7B model: Upgraded with WMT25-champion accuracy & fewer artifacts. 🌍 [--] languages + [--] Chinese dialects 🛠 Context-aware custom glossaries format-preserving output 💡 Perfect for edge + cloud deployments Models: https://modelscope.cn/collections/Tencent-Hunyuan/HY-MT15 https://modelscope.cn/collections/Tencent-Hunyuan/HY-MT15"
X Link 2025-12-30T10:08Z [----] followers, 17K engagements
"One of your New Year open-source gifts 🎨 is on its way"
X Link 2025-12-30T11:45Z [----] followers, 10.7K engagements
"🚀New on ModelScope: Qwen-Image-2512 is here 🎉 This version delivers a massive leap in realism and control: ✨ Photorealistic humans dramatically reduced AI look with accurate aging natural expressions and fine details (yes even individual hair strands) 🌿 Ultra-detailed nature fur foliage water and mist rendered with unprecedented texture fidelity 🔤 Reliable text & layout generates complex infographics timelines and multilingual slides with correct structure and legible text 🏆 Evaluated in 10k+ blind tests: now the strongest open-source image modelrivaling closed alternatives. 💡 Supports"
X Link 2025-12-31T09:53Z [----] followers, [----] engagements
"2026 is almost here Wishing you a very Happy New Year 🎉 What open-source models or product features are at the top of your wishlist for the coming year 🚀 Wed love to see if we can help make those wishes come true"
X Link 2025-12-31T12:28Z [----] followers, [----] engagements
"🌟Qwen-Image-2512 is now live on Model: Try: https://modelscope.ai/civision/imageGeneration/tab=default https://modelscope.ai/models/Qwen/Qwen-Image-2512/summary http://ModelScope.ai https://modelscope.ai/civision/imageGeneration/tab=default https://modelscope.ai/models/Qwen/Qwen-Image-2512/summary http://ModelScope.ai"
X Link 2025-12-31T17:37Z [----] followers, [----] engagements
"🚀 New on ModelScope: IQuest Research just dropped IQuest-Coder-V1 a new family of open-source code LLMs (7B to 40B) with 128K context GQA and two specialized variants: - Instruct: for coding assistance - Thinking: RL-tuned for agentic reasoning & error recovery - Plus: a novel Loop architecture (40B only) for efficient deployment via parameter sharing across [--] iterations. 🌟 All models support repo-level understanding via Code-Flow training learning from commit sequences and code evolution not just static snippets. 🔍 Benchmarks: - [----] on SWE-Bench Verified - [----] on LiveCodeBench v6"
X Link 2026-01-04T07:14Z [----] followers, 10.9K engagements
"Big news 🚀 now officially supports FREE LoRA training for Qwen-Image-2512 Z-Image-Turbo and Qwen-Image-Edit-2511. Time to build 🛠 🎨ModelScope Civision: https://modelscope.ai/civision/modelTraining http://ModelScope.ai https://modelscope.ai/civision/modelTraining http://ModelScope.ai https://modelscope.ai/civision/modelTraining http://ModelScope.ai https://modelscope.ai/civision/modelTraining http://ModelScope.ai"
X Link 2026-01-04T12:13Z [----] followers, 21.5K engagements
"Qwen-Image-2512-Fast is officially here ⚡ 🚀 Try the Demo on ModelScope now: Experience a massive 20x speedup generating high-quality images in just [--] to [--] steps Using CFG distillation weve achieved lightning-fast inference without compromising quality. https://modelscope.cn/studios/kelseye/Qwen-Image-2512-Turbo-LoRA-Demo/summary https://modelscope.cn/studios/kelseye/Qwen-Image-2512-Turbo-LoRA-Demo/summary https://modelscope.cn/studios/kelseye/Qwen-Image-2512-Turbo-LoRA-Demo/summary https://modelscope.cn/studios/kelseye/Qwen-Image-2512-Turbo-LoRA-Demo/summary"
X Link 2026-01-04T12:35Z [----] followers, 30.5K engagements
"@WolfyBlair Yes you'll get daily free quota (magicubes) and you can earn more by participating in community activities"
X Link 2026-01-05T02:26Z [----] followers, [---] engagements
"🚀 New on ModelScope: The Wuli Team just released a Turbo LoRA that makes Qwen-Image-2512 inference 20x faster⚡ without sacrificing quality. Key updates: ✅ 4-step turbo inference ✅ Enhanced texture & detail ✅ Native ComfyUI support Model: Demo: https://modelscope.cn/studios/kelseye/Qwen-Image-2512-Turbo-LoRA-Demo https://modelscope.ai/models/Wuli-Art/Qwen-Image-2512-Turbo-LoRA https://modelscope.ai/models/Wuli-Art/Qwen-Image-2512-Turbo-LoRA Qwen-Image-2512-Fast is officially here ⚡ 🚀 Try the Demo on ModelScope now: https://t.co/LRShwwbiNh Experience a massive 20x speedup generating"
X Link 2026-01-05T03:41Z [----] followers, 11.3K engagements
"🤖 Introducing InternVLA-A1 now fully open-sourced Many VLA models follow instructions well in static scenes but struggle in dynamic environments (conveyor belts rotating platforms multi-robot setups). Why They see the presentbut cant imagine the future. InternVLA-A1 solution: unify perception imagination and action in one model: ✅ Scene understanding: Image + text task parsing ✅ Task imagination: Predict future frames reason about dynamics ✅ Guided control: Execute actions steered by visual foresight Powered by InternData-A1 - Large-scale high-quality simulated dataset InternVLA-A1 stays"
X Link 2026-01-05T11:23Z [----] followers, 38.2K engagements
"🚀 Small model BIG agency Meet Youtu-LLM a native agentic 1.96B LLM that thinks plans and acts by itself no external frameworks needed. ✅ Outperforms 4B8B models on math coding & agent benchmarks ✅ Trained on 200B+ agentic reasoning trajectories ✅ 128K context open-weight & ready to run anywhere Think lightweight cant be smart Think again. 🤖 Model: 🔗 GitHub: 📄 Paper: https://www.modelscope.cn/papers/2512.24618 https://github.com/TencentCloudADP/youtu-tip/blob/master/youtu-llm https://modelscope.cn/collections/Tencent-YouTu-Research/Youtu-LLM https://www.modelscope.cn/papers/2512.24618"
X Link 2026-01-06T07:57Z [----] followers, 11.8K engagements
"🚀 TeleChat3-105B-A4.7B-Thinking is now open source A 105B sparse MoE model with fine-grained routing: - [---] experts only [--] activated per token (4.7B active params) - Trained end-to-end on domestic compute - Strong across code math agents writing check HumanEval-X (92.7%) & SWE-Bench (51%) It can even build a working shooter game 🕹 or a video site frontendfully autonomously. Also released: TeleChat3-36B-Thinking (dense 36B) optimized for reasoning & roleplay. ✅ Supports vLLM SGLang LLaMA-Factory ✅ Full Thinking mode with think./think reasoning traces ✅ Tuned for long-context and multi-turn"
X Link 2026-01-06T11:20Z [----] followers, 16.8K engagements
"The #EmbodiedAI data gap just got a lot smaller. 🚀 @AGIBOTofficial released the Genie Sim [---] datasetthe largest open-source sim dataset in the field 🤖 10000+ hours of high-quality sim demos 🎯 200+ tasks & 100k scenarios 📊 Multi-sensor: RGB-D Stereo & Kinematics Standardized in LeRobot format. Ready to train ✅ Dataset: ✅ GitHub: https://github.com/AgibotTech/genie_sim https://modelscope.cn/datasets/agibot_world/GenieSim3.0-Dataset https://github.com/AgibotTech/genie_sim https://modelscope.cn/datasets/agibot_world/GenieSim3.0-Dataset"
X Link 2026-01-07T07:00Z [----] followers, [----] engagements
"AI for all 🌍 Hosted by @alibaba_cloud the Alibaba Cloud AIGC Championship @ Milano Cortina [----] is officially live Create your Winter Olympics videos with the tools and resources powered by ModelScope as an official partner. ⛸❄ 🏆 Win Big 🎟 Milan [----] Tickets: A trip to the Winter Olympics 🏛 Legacy: Works archived in the Olympic Museum. 💰 Cash Prizes: Reward your creative storytelling. 🎬 Pro Tech: High-speed AI video via Wan [---]. Global Entrance Creating #ModelScope #AlibabaCloud #MilanoCortina2026 #Wan26 #Olympics https://modelscope.ai/civision/videoGeneration"
X Link 2026-01-07T08:54Z [----] followers, [----] engagements
"👏ComfyUI now supports Z-Image-Turbo LoRAs trained with ModelScope Civision Huge thanks to our amazing community members for the incredibly fast integration 👉 Everyone is welcome to train LoRAs for free on Civision and thanks to your valuable feedback training speed has been significantly improved 🚀 https://modelscope.ai/civision/modelTraining https://github.com/Comfy-Org/ComfyUI/pull/11805 https://github.com/Comfy-Org/ComfyUI/pull/11805 https://modelscope.ai/civision/modelTraining https://github.com/Comfy-Org/ComfyUI/pull/11805 https://github.com/Comfy-Org/ComfyUI/pull/11805"
X Link 2026-01-12T03:16Z [----] followers, 11.9K engagements
"GLM-Image is here Try it out and let us know what you think: https://x.com/ModelScope2022/status/2011262011997651194 🚀GLM-Image is now open-source a breakthrough in "Cognitive Generation." By combining Autoregressive models with Diffusion Decoders it finally solves the struggle of generating posters PPTs and knowledge-heavy infographics. 🌟 Highlights: ✅ #1 on CVTG-2K & LongText-Bench https://t.co/cF3z0VFvc9 https://x.com/ModelScope2022/status/2011262011997651194 🚀GLM-Image is now open-source a breakthrough in "Cognitive Generation." By combining Autoregressive models with Diffusion"
X Link 2026-01-14T02:40Z [----] followers, [----] engagements
"Step-Audio-R1.1 by @StepFun_ai just set a new SOTA on the Artificial Analysis Speech Reasoning leaderboard 🏆 It outperforms Grok Gemini and GPT-Realtime with a 96.4% accuracy rate. ✅ Native Audio Reasoning (End-to-End) ✅ Audio-native CoT (Chain of Thought) ✅ Real-time streaming inference ✅ FULLY OPEN SOURCE 🌟 Demo: 🤖 Model: https://modelscope.cn/models/stepfun-ai/Step-Audio-R1.1 https://modelscope.cn/studios/stepfun-ai/Step-Audio-R1 https://modelscope.cn/models/stepfun-ai/Step-Audio-R1.1 https://modelscope.cn/studios/stepfun-ai/Step-Audio-R1"
X Link 2026-01-15T06:32Z [----] followers, 15.6K engagements
"@SteveWarnerFL Hi would you mind sharing more about how you use it in your work Are you trying to modify a specific layer or add/remove one Wed love to better understand your situation"
X Link 2026-01-16T06:34Z [----] followers, [--] engagements
"Step3-VL-10B Demo: https://modelscope.cn/studios/stepfun-ai/step3-vl-10b https://modelscope.cn/studios/stepfun-ai/step3-vl-10b"
X Link 2026-01-19T11:34Z [----] followers, [----] engagements
"🚀 Introducing GLM-4.7-Flash the strongest 30B-class open model that balances performance & efficiency ✅ Top scores on AIME GPQA SWE-bench -Bench & more ✅ Optimized for agentic coding tool use and reasoning ✅ Runs locally with vLLM / SGLang / Transformers Perfect for developers who want Claude-level coding power at a fraction of the cost 💡 Try it now: 🤖 🔗 📄 🐙 https://github.com/zai-org/GLM-4.5 https://z.ai/blog/glm-4.7 https://chat.z.ai https://modelscope.cn/models/ZhipuAI/GLM-4.7-Flash https://github.com/zai-org/GLM-4.5 https://z.ai/blog/glm-4.7 https://chat.z.ai"
X Link 2026-01-19T15:32Z [----] followers, 11.3K engagements
"How reliable is your model's tool usage. REALLY 🧐 Inspired by @Kimi_Moonshot's K2 Vendor Verifier we are leveling up Agentic Evaluation with EvalScope. The standardizing Function Calling evaluation allows us to bridge the gap between "claims of support" and "RELIABLE execution" in the field. 🛠 📖 Docs: 💻 Code: 🔹 Standardized: Quantifies MoonshotAI K2 verification logic. 🔹 Comprehensive: Measures decision-making + JSON schema compliance. 🔹 Simple: Plug-and-play with your data or our benchmarks. Don't let your Agent go live without a proper check-up #LLM #AIAgents #FunctionCalling"
X Link 2026-01-22T06:32Z [----] followers, [----] engagements
"@YashasGunderia Hi you'll need to request to join the organization first and then you'll have access to the GPU resources. Here is the link: https://www.modelscope.ai/organization/xGPU-Explorers https://www.modelscope.ai/organization/xGPU-Explorers"
X Link 2026-01-24T05:43Z [----] followers, [---] engagements
"Weve prepared ModelScope and Tongyi swag for everyone attending the workshopready for a day full of great takeaways 🚀 Join Us in Seoul: AI Innovation Meets Creativity @AMD @Alibaba_Qwen @ModelScope2022 @alibaba_cloud AMDs AI Developer Meetup in Seoul (Dec 10) is filling FAST. As a key partner were bringing you the future of generative AI. 📅 Dec [--] 📍 Seoul 🎁 Free https://t.co/b5XcmEmbPf 🚀 Join Us in Seoul: AI Innovation Meets Creativity @AMD @Alibaba_Qwen @ModelScope2022 @alibaba_cloud AMDs AI Developer Meetup in Seoul (Dec 10) is filling FAST. As a key partner were bringing you the"
X Link 2025-12-10T01:28Z [----] followers, [----] engagements
"Qwen-Image-Layered-Control is now supported by ComfyUI Try it out: https://modelscope.ai/models/DiffSynth-Studio/Qwen-Image-Layered-Control/files DiffSynth-Studio has open-sourced Qwen-Image-Layered-Control 🚀 It enables precise text-controlled layer extractionperfect for poster decomposition and graphic design workflows. Control image layers like never before: ✅ Single-layer output via text prompts ✅ Native https://t.co/4QwOKqTFLs https://modelscope.ai/models/DiffSynth-Studio/Qwen-Image-Layered-Control/files DiffSynth-Studio has open-sourced Qwen-Image-Layered-Control 🚀 It enables precise"
X Link 2026-01-15T09:33Z [----] followers, 22.9K engagements
"Latency or intelligence Step-Audio-R1.1 ends the trade-off. Fei Tian Multimodal Large Model Researcher at StepFun introduces Mind-Paced Speakingenabling real-time voice AI that thinks while speaking. Fast responses. Deep reasoning. No compromise. https://x.com/i/broadcasts/1mrGmBabZavJy https://x.com/i/broadcasts/1mrGmBabZavJy"
X Link 2026-01-19T11:57Z [----] followers, [----] engagements
"Introducing AgentCPM-Report: An 8B on-device agent that rivals closed-source giants in deep research and professional report writing. The Performance: 🏆 Ranked #1 in Insight Depth across DeepResearch Bench Deep Consult and DeepResearch Gymoutperforming top-tier closed models. Highlights: ✅ 100% Private: Zero data leakage. Your sensitive data stays on your hardware. ✅ UltraRAG Stack: Efficient local indexing for private PDFs and docs. ✅ Production Ready: One-click Docker deployment. How does an 8B model hit SOTA 🧠 It breaks the "one-shot" generation bottleneck with a "Writing as Reasoning""
X Link 2026-01-20T12:18Z [----] followers, [----] engagements
"🚀 Chroma [---] is here and its open From @flashlabsdotai : the worlds first open-source end-to-end real-time speech-to-speech dialogue model with personalized voice cloning. ✨ 150ms end-to-end latency 🧬 High-fidelity voice cloning from just seconds of audio 📈 Speaker similarity (SIM) = [-----] +10.96% over human baseline 🧠 Strong reasoning in only 4B params 🔓 Fully open weights + code A true open alternative to closed commercial systems and already optimized with @lmsysorgs SGLang for even faster inference 🔗 Paper: 🤖 Model: 💻 Code: Were excited to see what the community builds with"
X Link 2026-01-21T16:07Z [----] followers, 14.3K engagements
"🌾 Meet Sinong (): The first heavy-duty LLM suite for Agriculture 🚜 🧬 Expertise: Crop breeding animal medicine smart farming & ag-economics. 🤖 Models: 8B & 32B parameters 📚 Stack: 4B+ tokens (240k+ papers 8k+ books patents) 🧠 Tech: Native CoT reasoning + Multi-Agent RAG for zero-hallucination scientific retrieval. 🚀 Open Source: https://modelscope.cn/models/NAULLM/Sinong1.0-32B https://modelscope.cn/models/NAULLM/Sinong1.0-8B https://modelscope.cn/models/NAULLM/Sinong1.0-32B https://modelscope.cn/models/NAULLM/Sinong1.0-8B"
X Link 2026-01-23T03:46Z [----] followers, [----] engagements
"@noctus91 Hi there This isn't actually an error. To use API inference you just need to bind your Alibaba Cloud account (no extra cost involved)"
X Link 2026-01-23T17:37Z [----] followers, [---] engagements
"Meet LingBot-VA: The future of robot learning is visual. 🤖🎥 Its an autoregressive diffusion framework that predicts future video frames and decodes actions simultaneously. Instead of just reacting it reasons: "I expect the world to look like this next so I should move like that." Key Highlights: 🧠 Infinite Memory: Unlike memoryless models that get stuck in loops LingBot-VA handles "recurrent states" (like opening the same box twice) with ease by remembering the full history. 🍳 Generalist Skills: From folding clothes to making breakfast and high-precision screw pickingits a true"
X Link 2026-01-30T02:47Z [----] followers, [----] engagements
"ChenkinNoob-XL v0.2 is here A high-fidelity SDXL-to-Rectified Flow conversion delivering vivid color range and superior lighting adherence. ✅ Performance: Eliminates "base model greyness"; Superior contrast/dark scene rendering; High stability across wide CFG ranges (3-6); Optimal convergence at 20-28 steps. ✅ Ecosystem: Native ComfyUI workflow included; ReForge-ready with built-in Flow support; LoRA-compatible via Pochi.toml (Refresh scripts). ✅ Architecture: Rectified Flow (RF) linear ODE trajectory; Trained on 47M samples using 8xH20 nodes; Leverages Cosine Optimal Transport and Protected"
X Link 2026-02-06T05:00Z [----] followers, [----] engagements
"VibeThinker-1.5B is here 🚀 and it flips the bigger = smarter myth on its head. ✅ Just 1.5B params ✅ Trained via novel Spectrum-to-Signal Principle (SSP) ✅ Beats models 400x larger (e.g. 671B DeepSeek-R1) on hard math benchmarks (AIME24/25 HMMT25) ✅ Matches 456B MiniMax-M1 and ties Mistrals Magistral-Medium on LiveCodeBench v6 ✅ Post-training cost: $8K (vs. $290K+ for others) 💡 SSP first encourages divergent exploration of solution paths then uses RL to converge on optimal strategies smarter training not bigger models. 🌍 Open-sourced to empower researchers & smaller teams whove been priced"
X Link 2025-11-13T13:00Z [----] followers, [----] engagements
"Huge update for Z-Image-Turbo-Fun-Controlnet-Union [---] 🚀 Key technical upgrades: ✅ New 1.9GB Lite model for low-VRAM & natural blending ✅ Fixed mask leakage in inpainting ✅ Multi-resolution dataset refactor (up to 1536px) ✅ 8-step distillation for crisp blur-free Turbo gen No more bright spot artifacts. Just pure control. 🤖 Model: https://modelscope.ai/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union-2.1 https://modelscope.ai/models/PAI/Z-Image-Turbo-Fun-Controlnet-Union-2.1"
X Link 2026-01-14T08:34Z [----] followers, 20.5K engagements
"Real-Qwen-Image-V2 is here 📸 Developed by wikeeyang this fine-tuned version of Qwen-Image-2512 is built for those who demand peak realism and production-grade quality. 🤖 Model: Key Highlights: ✨ Sharper details & enhanced realism ✨ Optimized for Asian facial aesthetics ✨ Superior text & image generation quality ✨ Highly compatible with LoRAs & custom workflows Tech Specs: ✅ CFG: [---] - [---] ✅ Steps: [--] - [--] ✅ Sampler: Euler / Simple (flexible) ✅ Model Shift: [---] - [---] https://www.modelscope.cn/models/wikeeyang/Real-Qwen-Image-V2 https://www.modelscope.cn/models/wikeeyang/Real-Qwen-Image-V2"
X Link 2026-01-18T05:01Z [----] followers, 28.4K engagements
"🚀 Meet STEP3-VL-10Bit delivers SOTA-level visual perception complex reasoning and human-aligned intelligenceredefining efficiency in open multimodal AI. ✅ Beats or matches models [----] larger (like GLM-4.6V Qwen3-VL even Gemini [---] Pro) ✅ Achieves SOTA on MMMU MathVision OCRBench ScreenSpot and more ✅ Trained on 1.2T tokens + 1400+ RL rounds (RLHF + RLVR) ✅ Supports PaCoRe: parallel collaborative reasoning (128K context) Despite its compact size it leads the 10B class in: STEM reasoning (94.43% on AIME [----] w/ PaCoRe) Visual perception (92.05 on MMBench) GUI understanding & OCR Spatial"
X Link 2026-01-19T11:32Z [----] followers, 217.6K engagements
"⚡🎨Tomorrow. This week is going to be intriguing"
X Link 2026-01-26T02:30Z [----] followers, 51.8K engagements
"DeepSeek-OCR-2 is now on ModelScope ✨ Meet DeepSeek-OCR-2 a vision-language model that reimagines document intelligence with human-like visual understanding. Key Highlights 🧠 Visual Causal Flow: Understands layout semantics and structurenot just pixels 📏 Dynamic Resolution: Processes documents at (0-6)768 + [----] with adaptive token generation ✍ Flexible Prompts: One model for multiple tasks: - Convert to markdown preserve tables & formatting - Parse the figure extract chart insights - Locate "text" pixel-level grounding - OCR this image 100+ language support ⚡ Production Ready:"
X Link 2026-01-27T06:26Z [----] followers, [----] engagements
"🚀 Meet Kimi K2.5 🌙 This is Kimis most intelligent and versatile model to date achieving SOTA performance across coding vision and agentic workflows. Model: Paper: Highlights: ✅ Native Multimodal Architecture: Seamlessly integrates vision & text. From screenshots to screen recordings K2.5 "sees" and reasons across images and videos to break the limits of text-only prompts. ✅ Agent Clusters: A world-first capability. K2.5 can autonomously spawn up to [---] "clones" to work in parallel on complex tasks slashing execution time by 4.5x. ✅ The Ultimate Office/Coding Assistant: * Office Pro: Masters"
X Link 2026-01-27T06:58Z [----] followers, [----] engagements
"The DiffSynth-Studio team also dropped Z-Image-i2L (Image-to-LoRA). 🎨✨ This model takes a single image as input and instantly outputs a custom LoRA tailored to that specific style or feature. Model: Demo: ✅ Zero Manual Effort: No captioning or training process required. ✅ One-Shot Magic: Get a style LoRA from just one reference image. ✅ Style Transfer: Seamlessly extract and apply unique features to new prompts. ✅ Lower Barrier: Personalized model creation is now accessible to everyone. https://modelscope.cn/studios/DiffSynth-Studio/Z-Image-i2L"
X Link 2026-01-27T16:02Z [----] followers, [----] engagements
"🚀 Meet LingBot-VLA: A pragmatic Vision-Language-Action model designed to bridge the gap between perception and execution in robotics. 🤖 ✅LingBot-VLA-4B: Lightweight & versatile. ✅LingBot-VLA-4B-Depth: Enhanced for high-precision spatial tasks. Highlights: 🧠 Powerful Core: Built on the Qwen2.5-VL-3B foundation mastering multi-tasking and dual-arm coordination across 9+ robot configs. 📈 Elite Performance: Outperforms competitors like [---] and GR00T in success rates (SR) on both GM-100 (Real-world) and RoboTwin [---] (Sim). 🔋 Hyper-Efficient: 1.52.8x faster training than existing VLA codebases"
X Link 2026-01-27T18:00Z [----] followers, [----] engagements
"Introducing LingBot-World: An open-source world simulator pushing the boundaries of video generation. 🚀 🌍 High-Fidelity: Realistic scientific & stylized. 🧠 Long-Term Memory: Minute-level consistency. ⚡ Real-Time: 1s latency at [--] FPS. 📜 Apache [---] Licensed. Model: Github: https://github.com/Robbyant/lingbot-world https://modelscope.cn/models/Robbyant/lingbot-world-base-cam https://github.com/Robbyant/lingbot-world https://modelscope.cn/models/Robbyant/lingbot-world-base-cam"
X Link 2026-01-28T17:30Z [----] followers, 29.3K engagements
"HunyuanImage-3.0 is here: A high-performance multimodal model for unified text-to-image generation and creative editing. 🚀🧵👇 🤖Model: 1/ The Model Suite 📦A versatile lineup for any workflow: ✅Base 3.0: The core T2I powerhouse. ✅3.0-Instruct: Advanced I2I and creative editing. ✅Instruct-Distil: Optimized for 8-step sampling and rapid deployment. 2/ Core Capabilities 🎨Beyond renderingits a creative partner. Features Unified Generation (T2I I2I and 3-image fusion) with built-in CoT reasoning and Prompt Enhancement to turn sparse ideas into professional descriptions. 3/ Technical Highlights"
X Link 2026-01-29T11:51Z [----] followers, [----] engagements
"PaddleOCR-VL-1.5 is open-sourced A 0.9B parameter multimodal model that masters the "curved & distorted" document challenge. 🚀 🔗 ModelScope: 🔗 Demo: Key Technical Breakthroughs: ✅Polygon-based Localization: No more rigid rectangles. It natively supports irregular boxes to fit warped tilted or screen-captured text perfectly. 📐 ✅Seal & Stamp Recognition: New dedicated capability to handle complex official document verification. 💮 ✅Cross-Page Logic: Automatically merges tables and identifies headings across physical page breakscrucial for full-document semantics. ✅Performance Giant:"
X Link 2026-01-30T05:52Z [----] followers, [----] engagements
"Tencent YouTu Research open-sourcedYoutu-VL-4B-Instruct a compact VLM redefining visual perception viaVLUAS. 👁⚡ 🌟Highlights: ✅All-in-One Vision: SOTA performance in detection segmentation depth estimation and pose estimationwithouttask-specific heads. ✅OCR & Reasoning: Exceptional at complex document parsing and multimodal math reasoning. ✅GUI Agent Ready: Optimized for world-understanding and interface navigation. ✅Efficiency: 4B parameters make it perfect for edge deployment and fast inference. 🔧Performance: Outperforms many larger models inOmniDocBenchandVision-Centrictasks while"
X Link 2026-01-30T11:00Z [----] followers, [----] engagements
"Ready to make this February legendary 🚀✨The Qwen-Image LoRA Training Competition is officially LIVE Weve teamed up with @Ali_TongyiLab and @Alibaba_Qwen to bring you some absolutely grand prizes:📱 iPhone [--] Pro Max (512GB) 🎮 Sony PS5 Pro 🛍 $800 Shopping Cards (or equivalent prizes) Stop scrolling and start training 🛠🎨Join here: #HappyQwensday #QwenImageLoRA https://modelscope.ai/active/qwenimagelora https://modelscope.ai/active/qwenimagelora"
X Link 2026-02-01T12:35Z [----] followers, [----] engagements
"Beyond-Reality-Z-Image [---] is here.🚀 Highlights: ✅Texture SOTA: Optimized high-frequency details for what might be the best out-of-the-box skin and fabric textures in the Z-model family. ✅Cinematic Lighting: Re-engineered color balance to achieve that authentic "film look" without heavy prompting. ✅Speed x Versatility: Generates at Turbo speeds (10-15 steps) but supports broader artistic styles and 2D/Anime aesthetics than the official release. ✅Efficiency: Optimized for low CFG (1.0) making it highly stable and responsive to LoRA integration. Available now on ModelScope: Recommended Stack:"
X Link 2026-02-02T13:03Z [----] followers, 16.4K engagements
"2/3 🛠Four specialist models were released in just one week each mastering a specific real-world challenge: ✅ DeepSeek-OCR2: Uses "Visual Causal Flow" to handle reading order in complex multi-column layouts. ✅ GLM-OCR (Z-org): An efficiency powerhouse using Multi-Token Prediction (MTP) to reduce costs by 90% while excelling at formulas. ✅ PaddleOCR-VL-1.5 (PaddlePaddle): Built for "Real5" messy conditionsperfect for curved pages and camera glare. ✅ Youtu-Parsing (Tencent): Delivers 22x faster inference and converts flowcharts directly into Mermaid code."
X Link 2026-02-03T09:18Z [----] followers, [----] engagements
"Here they come They are now at or alumni of these elite AI labs: Alibaba Tongyi Meta (FAIR/Superintelligence) Google DeepMind Tencent AI. Academia: SJTU OSU UChicago NUS WashU. https://x.com/i/broadcasts/1ZkJzZRdQNoJv https://x.com/i/broadcasts/1ZkJzZRdQNoJv"
X Link 2026-02-05T05:55Z [----] followers, [----] engagements
"@pranaysuyash http://ModelScope.ai http://ModelScope.ai"
X Link 2026-02-05T12:38Z [----] followers, [---] engagements
"1/4🧵 MS-Agent is Rebuilding A massive functional update to ModelScopes agent frameworkintroducing autonomous skill scaling and industrial-grade toolkits. ✅ Agent Skills: Skill Discovery via local/Hub repo_id matching; ✅ Multi-skill Collaboration via DAG-based planning; ✅ Progressive Analysis to optimize context window for small models. ✅ Robust Executor: Secure Docker/Sandbox execution; autonomous self-reflection and bug-fix cycles. Github: https://github.com/modelscope/ms-agent/tree/main/ms_agent/skill https://github.com/modelscope/ms-agent/tree/main/ms_agent/skill"
X Link 2026-02-06T02:34Z [----] followers, [---] engagements
"2/4 💻 Code Genesis is Now Live A SOTA end-to-end synthesis engine specializing in zero-to-one development over standard "vibe-coding." ✅ Capabilities: Industrial-grade Code Genesis for full-stack synthesis; Language Server integration for stability and clean code aesthetics. ✅ Technical Edge: Superior benchmarks in end-to-end logic flow; minimalist architecture for rapid project scaffolding. Repo: Tech Report: https://github.com/modelscope/ms-agent/tree/main/projects/code_genesis/TECH_REPORT.md https://github.com/modelscope/ms-agent/tree/main/projects/code_genesis"
X Link 2026-02-06T02:34Z [----] followers, [---] engagements
"🎬 Singularity Cinema is Breaking Benchmarks The GLCC-winning workshop for high-density long-form automated video production. ✅ Generation: LLM-driven storyboards; Manim/Remotion dual-engine animation; 5min+ cinematic output for Science Econ and History. ✅ Model Fusion: Seamlessly integrates Qwen-Image Sora edge-tts and nano-banana-pro; transforms static docs into dynamic media. Demo: https://github.com/modelscope/ms-agent/tree/main/projects/singularity_cinema https://github.com/modelscope/ms-agent/tree/main/projects/singularity_cinema"
X Link 2026-02-06T02:34Z [----] followers, [---] engagements
"4/4 🌍 ms-agent WebUI is Live A self-bootstrapped interface generated entirely by the Code Genesis engine. ✅ Usage: pip install ms-agent -U - ms-agent ui; instant GUI access for agent orchestration. ✅ Roadmap: DeepResearch capability launching soonSOTA autonomous retrieval and synthesis incoming. Full Framework: https://github.com/modelscope/ms-agent https://github.com/modelscope/ms-agent"
X Link 2026-02-06T02:34Z [----] followers, [---] engagements
"Thinking Machines Lab proved On-Policy Distillation slashes LLM training costs by 10x and we show you how to reproduce their research. Invest [--] minutes in this guideas we unpack the theory tech details experiment results and code to instantly transform your fine-tuning budget📚 Related Resources👇): ✅ Slash training compute by 10X. ✅ Achieve robust RL performance with zero forgetting. ✅ Get the ready-to-use ms-SWIFT + vLLM code for deployment. Related Resources - TML Blog: - (Open-source implementation for reproducing On-Policy Distillation) - On-Policy Distillation Documentation: - Example"
X Link 2025-10-28T13:09Z [----] followers, [----] engagements
"Stepfun open-sourcedStep-3.5-Flash a powerhouse model specifically architected for high-speed reasoning and complex Agentic workflows. 🚀 Model: Key Technical Specs: ✅Sparse MoE Architecture: 196B total params but only11B activeper token. SOTA efficiency. ✅MTP-3 (Multi-Token Prediction): It predicts3 tokens at once hitting a blistering350 TPSfor code-heavy tasks. ⚡ ✅Hybrid Attention (SWA + Full): A 3:1 mix that masters256K contextwindows while keeping compute costs low. ✅Parallel Thinking: Massively boosted performance for multi-step reasoning and deep search. Why Devs should care: - Built"
X Link 2026-02-02T06:52Z [----] followers, 34K engagements
"How do Agents "Grow from Experience" A dialogue on Key Issues and Technical Paths of Self-Evolution AI AMA Deep Roundtable Vol.1 [----] was about Reasoning. [----] is the "Mirror Moment"Agents are shifting from static tools to digital lifeforms that learn iterate and evolve from every failure. 🚀 https://twitter.com/i/web/status/2018712297829732601 https://twitter.com/i/web/status/2018712297829732601"
X Link 2026-02-03T15:44Z [----] followers, [----] engagements
"Meetthe1stAuthorsbehindthemostinfluentialpapers:earlyexperienceDREAMGYMAgentEvolverSPICER-ZeroEvo-MemoryandSEAgent. TheyarenowatoralumnioftheseeliteAIlabs: 🔹AlibabaTongyiLabMetaFAIRMetaSuperintelligenceLabsGoogleDeepMindTencentAISeattleLab 🔹SJTUOSUUChicagoNUSWashU 6leadresearchers1stage.Unrivaledtechnicaldepth.🧠 https://twitter.com/i/web/status/2018712300744847796 https://twitter.com/i/web/status/2018712300744847796"
X Link 2026-02-03T15:44Z [----] followers, [---] engagements
"MiniCPM-o [---] brings true "Omni" capabilities (Vision + Audio + Text) into a single 9B end-to-end architecture (Qwen3-8B + SigLip2 + Whisper + CosyVoice2).🚀 ✅ Vision SOTA: OpenCompass score of 77.6surpassing GPT-4o and Gemini [---] Pro in VLM tasks with just 9B params. ✅ Real-time Full Duplex: It "sees" "hears" and "speaks" simultaneously. Supports active interaction (it can interrupt or initiate conversation based on live video). ✅ World-Class OCR: Beats DeepSeek-OCR [--] and GPT-4o in document parsing (OmniDocBench leader). ✅ Voice Cloning: Clone any voice with just a short reference"
X Link 2026-02-04T05:51Z [----] followers, [----] engagements
"🚀 Meet Intern-S1-Pro: A massive 1T parameter MoE model for Multimodal Science Reasoning ✅ [---] Experts (22B active) ✅ SOTA in AI4Science (Chemistry Materials Bio) ✅ FoPE + Time-series modeling (up to [--] points) ✅ Native "Thinking Mode" support Open-source science just leveled up. 🧪💻 Model: Github: https://github.com/InternLM/Intern-S1 https://www.modelscope.cn/models/Shanghai_AI_Laboratory/Intern-S1-Pro https://github.com/InternLM/Intern-S1 https://www.modelscope.cn/models/Shanghai_AI_Laboratory/Intern-S1-Pro"
X Link 2026-02-04T15:20Z [----] followers, [----] engagements
"From a local designer in Chengdu to a global AI sensation with 117k+ downloadsDaniels journey shows whats possible when open-source meets imagination. Now its your turn. Join the Qwen-Image LoRA Competition and start building your own legacy. The next top creator could be you. https://modelscope.ai/active/qwenimagelora LoRA Training Contest: Win iPhone/PS5 and other prizes Gaussian splash LoRA creation process shared. Competition information: https://t.co/MEEZHQcr83 Training address: https://t.co/uxwtMlJzoP @Ali_TongyiLab @ModelScope2022 #HappyQwensday #QwenImageLora https://t.co/5igMVZUO5i"
X Link 2026-02-06T03:20Z [----] followers, [----] engagements
"StepFun open-sourced Step-3.5-Flashfeaturing powerful reasoning and Agent intelligence. Speeds up to [---] TPS Competes with closed-source models in Agent and Math tasks Excels at complex long-chain reasoning. Dont miss out https://x.com/i/broadcasts/1OyKAjZgZPaGb https://x.com/i/broadcasts/1OyKAjZgZPaGb"
X Link 2026-02-06T12:05Z [----] followers, [----] engagements
"YOLO26 is here🚀 The most advanced and deployment-ready vision model from @ultralytics redefining real-time edge computing performance. ✅ Performance: 43% CPU inference speed boost vs YOLO11; Real-time performance on low-power hardware; [--] scalable sizes (n/s/m/l/x) for optimized speed-accuracy trade-offs. ✅ Ecosystem: Unified framework for Detection Segmentation Pose OBB and Tracking; Simplified export for seamless hardware integration. ✅ Logic/Architecture: Removed DFL for enhanced hardware compatibility; End-to-end NMS-free inference; ProgLoss + STAL for small target precision; MuSGD"
X Link 2026-02-09T08:30Z [----] followers, [----] engagements
"2/4 Key Technical Highlights: - Architecture: Native single-stream design integrating generation editing and segmentation. - Audio: First model to unify Speech Audio and Music generation via 12.5Hz continuous representation. - Video: Time-Interleaved VideoRoPE for enhanced temporal modeling. https://twitter.com/i/web/status/2021580867009671592 https://twitter.com/i/web/status/2021580867009671592"
X Link 2026-02-11T13:43Z [----] followers, [---] engagements
"🐿 Meet Sirchmunk: The agentic search engine that operates directly on your raw data. Turn files into evolving intelligence in real-time Build information retrieval pipelines that require ZERO vector databases ZERO pre-indexing and ZERO chunking: 🎯 Embedding-Free Retrieval Drop files in and search instantly. No ETL pipelines no vector silos and no setup hours. 🎲 Monte Carlo Evidence Sampling It doesn't "read" full docsit explores focuses and synthesizes. 500-page manual Only the most relevant snippets hit the LLM. 🧠 Self-Evolving Knowledge Every search builds reusable knowledge clusters."
X Link 2026-02-13T13:39Z [----] followers, [---] engagements
"LongCat-Flash-LiteisNow Live Meituan's 68.5B MoE model pioneering "Embedding Expansion" for leading Agent and Coding performance. ✅Performance: 2.9B4.5B activated params; 500-700 tokens/s throughput; 256K context via YARN; SOTA SWE-Bench (54.4%) and -Bench scores. ✅Ecosystem: Optimized SGLang-FluentLLM engine with custom CUDA kernels; Open API with 50M free daily tokens. ✅Architecture: N-gram embedding layer for localized semantic precision; Hash-collision mitigation via sub-table decomposition; Triple-step Speculative Decoding synergy. 🤖 🔧 📄 Tech"
X Link 2026-02-08T06:27Z [----] followers, [----] engagements
"🚀 GLM-4.6V is live @Zai_org 's new open-source multimodal models just dropped on ModelScope and we provide a FREE API for immediate access❗ GLM-4.6V (106B) for cloud & high-performance workloads GLM-4.6V-Flash (9B) lightweight fast great for local inference 🔥 Key dev perks: ✅ Native multimodal tool calling pass images/docs directly as function args no OCR detour ✅ 128K context handles 150-page docs or hour-long videos in one go ✅ Visual Action pipeline powers real multimodal agents (e.g. find this outfit online returns structured shopping list) ✅ 50% cheaper than GLM-4.5V $1/million input"
X Link 2025-12-08T11:43Z [----] followers, [----] engagements
"Nanbeige just dropped Nanbeige4.1-3B a tiny powerhouse that outclasses models 10x its size.🚀 ✅ 256k Context Window ✅ "Deep Search" Agent capabilities built-in ✅ Two-stage Code RL (Correctness ➡ Efficiency) ✅ Beats Qwen3-32B on key benchmarks The era of high-performance low-cost local deployment is here. 🤖: https://modelscope.cn/models/nanbeige/Nanbeige4.1-3B https://modelscope.cn/models/nanbeige/Nanbeige4.1-3B"
X Link 2026-02-14T11:11Z [----] followers, 10.9K engagements
"Qwen3-ASR & Qwen3-ForcedAligner are now open-source 🎙🚀 High-performance streaming-ready speech AI for [--] languages and dialects. 🔗: Demo: 🤖The Models ✅ ASR-1.7B: Open-source SOTA. Outperforms Whisper-v3; rivals GPT-4o & Gemini. ✅ ASR-0.6B: Insane efficiency92ms latency; transcribes 2000s of audio in 1s. ✅ ForcedAligner-0.6B: Superior timestamp accuracy across [--] languages. 🌟Key Features ✅Universal: Supports [--] languages + [--] Chinese dialects. ✅Robust: Expertly handles singing heavy BGM and low SNR. ✅Unified: Single model for both streaming & offline (up to 20min audio). ⚡Performance &"
X Link 2026-01-29T12:55Z [----] followers, [----] engagements
"🚀 ACE-Step v1.5 is out: an open-source music foundation model that runs locally on consumer GPUs (4GB VRAM) and generates full songs in 2s (A100) or 10s (RTX 3090). ✅ Beats most commercial models in quality ✅ Train a personalized LoRA from just a few tracks ✅ Built on a novel LM+DiT hybrid architecture with internal RLno external reward bias ✅ Supports 50+ languages + editing (covers repainting vocalBGM) 🤖 https://modelscope.cn/models/ACE-Step/Ace-Step1.5 https://modelscope.cn/models/ACE-Step/Ace-Step1.5"
X Link 2026-02-04T09:19Z [----] followers, [----] engagements
"Introducing Singularity Cinema (MS-Agent framework)🚀an LLM-powered tool that transforms documents or prompts into storyboards generating stunning 5-minute+ videos via its Manim/Remotion dual-engine. By fusing nano-banana-pro🍌 Qwen-Image🎨 Sora and edge-tts it delivers high-quality information-dense contentperfect for science economics and history creators looking to scale their social media presence. 🔗: 🎦Prompt: Convert xx/deploy.md into a short video in a blue-themed style making sure to use the important images from the document. The short video must be in English. 🎬Output Video👇"
X Link 2026-02-05T16:08Z [----] followers, [----] engagements
"Meet X-Coder 🚀IIGroup drops a powerhouse suite for code LLMs ✅ X-Coder-SFT-376k: Massive synthetic dataset via GPT-o3-mini/DeepSeek-R1. ✅ X-Coder-RL-40k: Targeted RL alignment for logic & verification. ✅ X-Coder-8B Models: SOTA reasoning built on Qwen3. Experience the power of fully synthetic code evolution. 💻🔥 📊 Models & Dataset: #LLM #OpenCode #Qwen3 #DeepSeek #Coding https://modelscope.cn/models/IIGroup/X-Coder-SFT-Qwen3-8B https://modelscope.cn/datasets/IIGroup/X-Coder-RL-40k https://modelscope.cn/datasets/IIGroup/X-Coder-SFT-376k"
X Link 2026-02-09T06:25Z [----] followers, [----] engagements
"🚀Introducing Z-Image the full non-distilled version of Z-Image for high-quality customizable image generation. ✅ Model: ✅ Demo: ✨ Capabilities: Photorealistic + diverse art styles (anime illustration digital art) Full CFG & precise negative prompts for reliable prompt following Multi-subject generation with distinct identitiesno "clone faces" Community-friendly: ready for LoRA ControlNet and other fine-tuning workflows 🔥 Highlights: Full architecture (vs. distilled Turbo): better diversity & training stability ideal for customization High output variance across seeds = less homogenization"
X Link 2026-01-27T15:57Z [----] followers, 59.4K engagements
"How 0.9B Models Are Outperforming the Giants: A Comparison of Four New Chinese VLM-OCR Models 1/3 🏆 Early [----] has brought a massive surge of open-source models from DeepSeek Z-org PaddlePaddle and Tencent YouTu. A tiny 0.9B parameter model GLM-OCR topped the global OmniDocBench v1.5 leaderboard outscoring giants like Gemini-3 Pro and Qwen3-VL. We are moving away from rigid "detect and recognize" pipelines toward Vision-Language Models (VLMs) that actually "read" and understand document structures like humans do. https://twitter.com/i/web/status/2018615132134551565"
X Link 2026-02-03T09:18Z [----] followers, 30K engagements
"Z-Image-Fun-Controlnet-Union & Lora-Distill are now open-source 🚀Expanding the Z-Image ecosystem with ultra-fast inference and high-precision spatial control. ✅ Performance: Achieve 8-step inference (CFG: 1.0) via Distill LoRA and professional-grade multi-condition control (Canny Depth Pose Inpaint) across 15+ layer blocks. ✅ Ecosystem: Full cross-compatibility between Distill LoRA and Union ControlNet; includes "Lite" versions for low-VRAM edge devices and seamless natural blending. ✅ Unified Control: Support for 7+ spatial conditions in a single model with adjustable control_context_scale"
X Link 2026-02-05T09:17Z [----] followers, 10.9K engagements
"Z-Image-Fun-Lora-Distill is here A high-efficiency Distill LoRA enabling CFG-free high-fidelity generation for Z-Image in just 4-8 steps. ✅ 4-step and 8-step variants; CFG-free inference (CFG=1.0); [----] update delivers superior skin texture and color consistency compared to previous iterations. ✅ Native ComfyUI weights available; full integration with the VideoX-Fun framework; compatible with Z-Image ControlNet Union [---] and third-party LoRAs. ✅ Dual-distillation of both sampling steps and CFG; trained from scratch without Z-Image-Turbo weights; optimized for rapid generation and derivative"
X Link 2026-02-11T02:33Z [----] followers, [----] engagements
"🚀1/4 Introducing Ming-flash-omni-2.0: An open-source Omni-MLLM breakthrough. Based on the Ling-2.0 MoE architecture (100B total / 6B active) it unifies vision speech and image editing into a single model matching specialist performance across all modalities. 🧵👇 🤖 ModelScope: 📦 GitHub: https://github.com/inclusionAI/Ming https://www.modelscope.cn/models/inclusionAI/Ming-flash-omni-2.0 https://github.com/inclusionAI/Ming https://www.modelscope.cn/models/inclusionAI/Ming-flash-omni-2.0"
X Link 2026-02-11T13:43Z [----] followers, [----] engagements
"🚀 GLM-5 is now open-sourceand ranked #4 globally #1 among open models on Artificial Analysis Beyond "vibe coding" it delivers Agentic Engineering: autonomously architects debugs and ships production systems. Coding performance rivals Claude Opus [---] (SWE-bench: [----] 🔓 SOTA) while dominating long-horizon agent tasks (BrowseComp MCP-Atlas 🔝). 744B params MIT License 👉 ModelScope: https://modelscope.cn/models/ZhipuAI/GLM-5 https://modelscope.cn/models/ZhipuAI/GLM-5"
X Link 2026-02-12T03:43Z [----] followers, [----] engagements
"Love seeing what the MiniMax team shipped with M2.5 🚀 Real-world productivity strong agentic workflows and impressive efficiency numbers this is serious engineering. SOTA across coding search and tool use while getting faster is a big deal. 🔥 Excited to share that the model will be coming to ModelScope soon for the open-source community. Stay tuned 🌍✨ Introducing M2.5 an open-source frontier model designed for real-world productivity. - SOTA performance at coding (SWE-Bench Verified 80.2%) search (BrowseComp 76.3%) agentic tool-calling (BFCL 76.8%) & office work. - Optimized for efficient"
X Link 2026-02-12T17:09Z [----] followers, [----] engagements
"🚀 Ring-2.5-1T is open-source The first hybrid linear 1T thinking model. 📊 Gold Tier reasoning IMO25 (35/42) & CMO25 (105/126) scores. 🛠 Agent-native compatible with Claude Code OpenClaw and SGLang. ⚡Hybrid Linear breakthrough 10x lower memory & 3x throughput for long context. 🧠 1T MoE structure with 63B active parameters (1:7 MLA + Lightning Linear). 💻 Long-Horizon mastery autonomously built a 32-bit OS from scratch in [--] hours. Model: GitHub: https://github.com/inclusionAI/Ring-V2.5 https://modelscope.cn/models/inclusionAI/Ring-2.5-1T https://github.com/inclusionAI/Ring-V2.5"
X Link 2026-02-13T08:17Z [----] followers, 12.9K engagements
"🚀 FantasyWorld is now open source Recently accepted to ICLR [----] and ranking #1 on the WorldScore leaderboard this framework offers a unified approach to video and 3D generation. ✨Key Technical Highlights: ✅Direct 3D Inference: Input an image and camera trajectory to get frame-consistent video dense depth maps and point clouds instantly. ✅Dual-Branch IRG Architecture: An "Imagination" branch handles visual appearance while a "Geometry" branch ensures spatial accuracy. ✅Wan [---] Integration: The released 14B models leverage the latest Wan-DiT architectures for superior temporal consistency and"
X Link 2026-02-13T11:00Z [----] followers, [----] engagements
"🐿 Meet Sirchmunk: The agentic search engine that operates directly on your raw data. Turn files into evolving intelligence in real-time Build information retrieval pipelines that require ZERO vector databases ZERO pre-indexing and ZERO chunking: 🎯 Embedding-Free Retrieval Drop files in and search instantly. No ETL pipelines no vector silos and no setup hours. 🎲 Monte Carlo Evidence Sampling It doesn't "read" full docsit explores focuses and synthesizes. 500-page manual Only the most relevant snippets hit the LLM. 🧠 Self-Evolving Knowledge Every search builds reusable knowledge clusters."
X Link 2026-02-13T18:00Z [----] followers, [----] engagements
"Introducing FireRed-Image-Edit-1.0 from FireRedTeam 🚀 Its officially the new SOTA for general image editing. ✅ Better than Closed-Source: Outperforms Nano-Banana & Seedream4.0 on GEdit benchmarks. ✅ Native Evolution: Built from T2I foundations not just a "patch" on existing models. ✅ Style Mastery: Scored a record-breaking 4.97/5.0 in style transfer. ✅ High-Fidelity Text: Keeps original font styles perfectly. ✅ Virtual Try-on: Native support for multi-image joint editing. ✅ Bilingual: Native support for both English & Chinese prompts. Apache [---] license. Local deployment ready. 🤖 Model: 🎠"
X Link 2026-02-14T15:00Z [----] followers, 53.9K engagements
Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing