A Nebius mediu o impacto do NUMA binding num pretraining do DeepSeek V3
192 GPUs B300 rodando com Megatron Bridge no Managed Service for Kubernetes.
The same prompt word is given to the newly launched DeepSeek V4.1 Flash 1017 and Opus 5.5 to generate lyrics MV🔥
✅ DeepSeek (above): Cost 1.93 yuan, takes 32 minutes ✅ Opus (below): The cost is US$5.04 ≈ 36 yuan, which is nearly 19 times more expensive and takes 25 minutes. Key point: Both times were generated once, without modifying any line of code! Pictures, animations, and music points are all coded independently by the model.
DeepSeek V4.1 Flash just got a loud speed demo: ship a playable Tetris in about 10 seconds.
Isoquant posted a side-by-side "Tetris Battle" reel. Left: Claude Opus 5.5 on Anthropic's API with default reasoning. Right: DeepSeek V4.1 Flash 1017 Fast through Isoquant with reasoning
The KiiiKiii girls have thoughts on which frontier ai models to use, love that for them.
It won't be long before haum or jiyu have an opinion on which compressed/quantized variant of deepseek v4.1 1017 is optimal for local inference smart girls!
You have reached the end of the archive
All of deepseek