✨
AI Summary
- DeepSeek v3 launched as best open-weights model (LM Arena score 1319), making large model serving the bottleneck, requiring specialized inference infrastructure like Baseten's H200 clusters
- Chinese labs releasing 400B+ parameter models (Hunyuan-Large, MiniMax-Text) alongside DeepSeek v3, creating supply of high-quality open weights but infrastructure challenges
- Mission-critical inference requires deep optimization, collaboration with model builders, and hardware investment to serve state-of-the-art models in production