[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.]

@DFinsterwalder Avatar @DFinsterwalder David Finsterwalder | eu/acc

David Finsterwalder | eu/acc posts on X about open ai, $2413t, instead of, make a the most. They currently have XXXXX followers and XXX posts still getting attention that total XXXXXXX engagements in the last XX hours.

Engagements: XXXXXXX #

Engagements Line Chart

Mentions: XX #

Mentions Line Chart

Followers: XXXXX #

Followers Line Chart

CreatorRank: XXXXXXX #

CreatorRank Line Chart

Social Influence #

Social category influence technology brands #7176 stocks #1728 gaming finance

Social topic influence open ai #1551, $2413t #15, instead of #1078, make a, first ever, hosting, life of, loaded, gpu, minecraft

Top Social Posts #

Top posts by engagements in the last XX hours

"@CalebJohn24 I dont have a spark. Values are from this benchmark:"
X Link @DFinsterwalder 2025-10-15T07:56Z 1172 followers, XXX engagements

"@ppindia18 Grok is referencing the problematic ollama benchmark. Those values are not realistic. Spark gets XX t/s with GPT-OSS on llama.cpp and only XX t/s with ollama. Dont trust those values"
X Link @DFinsterwalder 2025-10-15T15:37Z 1172 followers, XX engagements

"It also wouldn't have been that easy without @OpenAI's excellent Codex. So while it would still have been unrealistic for my untechnical Archaeologist friend to build this completely himself we are getting closer to software-on-demand"
X Link @DFinsterwalder 2025-10-15T12:41Z 1176 followers, 2065 engagements

"@anthonyronning_ Both are MXFP4 via llama.cpp. I used llama.cpp integration from LM Studio though and the benchmark used ollama. Unfortunately there is no support for MLX for MXFP4 at this point. Usually MLX is even 10-20% faster"
X Link @DFinsterwalder 2025-10-15T07:48Z 1177 followers, XXX engagements

"@__phxlo @OpenAI No. Just instructions. But I have to say that I understand 3D graphics fairly well and in some case gave somewhat technical instructions. But if you are willing to learn some technical stuff just give it a try"
X Link @DFinsterwalder 2025-10-15T19:55Z 1177 followers, XX engagements

"Important info. The issue in that benchmark seems to be ollama. Native llama.cpp works much better. Not sure how ollama can fail so hard to wrap llama.cpp. The lesson: Dont use ollama. Espacially not for benchmarks"
X Link @DFinsterwalder 2025-10-15T08:07Z 1176 followers, 18.9K engagements

"DGX Spark GPT OSS 120B: XXXXX tok/sec M3 max GPT OSS 120B: XXXXX tok/sec This is very bad. Can't be explained from bad ollama performance and slower memory speed (273GB/s vs 400GB/s) alone"
X Link @DFinsterwalder 2025-10-14T12:09Z 1177 followers, 108.6K engagements

"@F_AI_Mouse @MaziyarPanahi Yeah they used ollama in that test which is using llama.cpp as well but consistently gives worse performance than original llama.cpp. This looks more reasonable"
X Link @DFinsterwalder 2025-10-14T18:54Z 1159 followers, XX engagements

"@Sohailm25 The super bad performance seems to be an ollama bug. But still M3 max and M4 max have higher GB/s bandwidth and should be faster. - DGX spark: XXX GB/s - M3 max: XXX GB/s - M4 max: XXX GB/s"
X Link @DFinsterwalder 2025-10-15T15:29Z 1174 followers, XX engagements

"@ggerganov Ollama is building a frontend and web hosting now. They cant be bothered anymore with something pedestrian like model implementations. Thats why they also ship models made for tool calls (Gwen coder) with broken tool calls wasting hours of the life of their once trusted users"
X Link @DFinsterwalder 2025-08-06T14:09Z 1089 followers, 5070 engagements

"@EHuanglu I dont think its disturbing at all for many (certainly for camera people though). But on the contrary: 100M can soon make XX great visual effects movies instead of X. People on streaming platforms would love more content"
X Link @DFinsterwalder 2025-10-15T08:41Z 1150 followers, XX engagements

"@levidehaan I know. I hate ollama. Its not my benchmark for the spark though. I used lm studio (which has a good llama.cpp) integration"
X Link @DFinsterwalder 2025-10-15T07:53Z 1173 followers, XXX engagements

"@walter4096 @MaziyarPanahi I don't know. It's not my benchmark values for DGX spark"
X Link @DFinsterwalder 2025-10-14T19:03Z 1158 followers, XX engagements

"@nonlinear_james Not really. I paid over 6000 for it. Currently M4 max with 128Gb RAM and 4TB is around 6000$ currently. The M4 max has 546GB/s memory per second bandwidth compared to sparks XXX GB/s though"
X Link @DFinsterwalder 2025-10-15T06:22Z 1172 followers, XXX engagements

"@Juicecountyeth Its not my benchmark for the spark. Values are from here:"
X Link @DFinsterwalder 2025-10-15T07:57Z 1173 followers, XXX engagements

"@aussetg It actually could be a benchmark issue. Others are reporting faster speeds using llama.cpp directly instead of ollama"
X Link @DFinsterwalder 2025-10-14T19:01Z 1172 followers, 1657 engagements

"@KaidenLarsen_ @OpenAI I didnt ask but I guess eventually it might have figured it out. I should just try with the screenshot and ask whats wrong"
X Link @DFinsterwalder 2025-10-15T17:43Z 1159 followers, X engagements

"@ivanfioravanti The issue in that benchmark seems to be ollama:"
X Link @DFinsterwalder 2025-10-15T09:47Z 1177 followers, XXX engagements

"@MaziyarPanahi So far the performance looks fairly bad. Four times slower on GPT-OSS 120B than M3 Max"
X Link @DFinsterwalder 2025-10-14T12:11Z 1174 followers, 2213 engagements

"@anoop_331 Right. Its really an ollama problem"
X Link @DFinsterwalder 2025-10-15T09:50Z 1172 followers, XX engagements

"@BadAINoData Yeah indeed. I also just saw native llama.cpp benchmark. Ollama results are horrible. The worst part is that they just wrap llama.cpp and it shouldnt be THAT bad. Incredible how much they fail"
X Link @DFinsterwalder 2025-10-15T08:01Z 1172 followers, XX engagements

"@gopinath9629 Hmm its XXX% loaded on GPU and running via Metal (using llama.cpp runtime from LM Studio). I don't think that ik_llama would make a difference here but I will try"
X Link @DFinsterwalder 2025-10-14T12:45Z 1176 followers, 2216 engagements

"It probably also wouldn't have been possible to do without all my prior knowledge in 3D graphics. For example we had a bug from georeferenced data where everything looked like Minecraft blocks. I instantly knew this was a floating-point precision issue from large coordinates"
X Link @DFinsterwalder 2025-10-15T12:41Z 1177 followers, 2554 engagements

"@anoop_331 Why is it not a fair comparison Both are running llama.cpp as runtime. If ollama didnt mess up it shouldnt make a difference (although ollama might have messed things up)"
X Link @DFinsterwalder 2025-10-15T07:51Z 1177 followers, XXX engagements

"@OpenAI I forgot to mention: Not only didnt I type a single line of code. I didnt type at all. I used MacWhisper to transcribe my instructions"
X Link @DFinsterwalder 2025-10-15T18:05Z 1177 followers, 1312 engagements

"In general the AMD Strix looks like a more competitive product price wise. But its not truly unified memory and the iGPU can only be assigned 96GB max on a 128GB system and the bandwidth is still lower. Theoretical max is 256Gb/s but that is much lower 120Gb/s or for cheap devices as they save money on the ram (exactly where the main bottle neck lies). Even the crippled bandwidth (276GB/s) of the GH10 is still much better than a low cost Strix"
X Link @DFinsterwalder 2025-10-15T10:16Z 1173 followers, XX engagements