[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]  Rohan Paul [@rohanpaul_ai](/creator/twitter/rohanpaul_ai) on x 73.9K followers Created: 2025-06-30 03:55:10 UTC π¨ CHINAβS BIGGEST PUBLIC AI DROP SINCE DEEPSEEK @Baidu_Inc open source Ernie, XX multimodal MoE variants π₯ Surpasses DeepSeek-V3-671B-A37B-Base on XX out of XX benchmarks π All weights and code released under the commercially friendly Apache XXX license (available on @huggingface ) thinking mode and non-thinking modes available π The 21B-A3B model beats Qwen3-30B on math and reasoning despite using XX% fewer parameters π§© XX released variants range from 0.3B dense to 424B total parameters. Only 47B or 3B stay active params, thanks to mixture-of-experts routing π A heterogeneous MoE layout sends text and image tokens to separate expert pools while shared experts learn cross-modal links, so the two media strengthen rather than hinder each other β‘ Intra-node expert parallelism, FP8 precision and fine-grained recomputation lift pre-training to XX% model FLOPs utilization on the biggest run, showing strong hardware efficiency πΌοΈ Vision-language versions add thinking and non-thinking modes, topping MathVista, MMMU and document-chart tasks while keeping strong perception skills π οΈ ERNIEKit offers LoRA, DPO, UPO and quantization for fine-tuning, and FastDeploy serves low-bit multi-machine inference with a single command βοΈ Multi-expert parallel collaboration plus 4-bit and 2-bit lossless quantization cut inference cost without hurting accuracy  XXXXXX engagements  **Related Topics** [coins ai](/topic/coins-ai) [Post Link](https://x.com/rohanpaul_ai/status/1939533142445953078)
[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]
Rohan Paul @rohanpaul_ai on x 73.9K followers
Created: 2025-06-30 03:55:10 UTC
π¨ CHINAβS BIGGEST PUBLIC AI DROP SINCE DEEPSEEK
@Baidu_Inc open source Ernie, XX multimodal MoE variants
π₯ Surpasses DeepSeek-V3-671B-A37B-Base on XX out of XX benchmarks
π All weights and code released under the commercially friendly Apache XXX license (available on @huggingface )
thinking mode and non-thinking modes available
π The 21B-A3B model beats Qwen3-30B on math and reasoning despite using XX% fewer parameters
π§© XX released variants range from 0.3B dense to 424B total parameters. Only 47B or 3B stay active params, thanks to mixture-of-experts routing
π A heterogeneous MoE layout sends text and image tokens to separate expert pools while shared experts learn cross-modal links, so the two media strengthen rather than hinder each other
β‘ Intra-node expert parallelism, FP8 precision and fine-grained recomputation lift pre-training to XX% model FLOPs utilization on the biggest run, showing strong hardware efficiency
πΌοΈ Vision-language versions add thinking and non-thinking modes, topping MathVista, MMMU and document-chart tasks while keeping strong perception skills
π οΈ ERNIEKit offers LoRA, DPO, UPO and quantization for fine-tuning, and FastDeploy serves low-bit multi-machine inference with a single command
βοΈ Multi-expert parallel collaboration plus 4-bit and 2-bit lossless quantization cut inference cost without hurting accuracy
XXXXXX engagements
Related Topics coins ai
/post/tweet::1939533142445953078