Kimi K3.1 Pro model weights are out

580 points · 165 comments on HN · read original →

Points and comments are a snapshot, not live.

Moonshot AI releases open-weight Kimi K3.1 Pro, a 2.8T-parameter multimodal agentic model.

Kimi K3.1 Pro is a Mixture-of-Experts model with 2.8T total parameters (104B activated), 1M-token context, and native multimodal support (text, image). Post-trained with joint reinforcement learning, it improves coding, 3D creation, financial research, and agentic tasks. Benchmarks show competitive performance against GPT-6 Astra and Claude Opus 5.5 in coding and agentic tasks. It uses KDA and Gated DSA attention. Deployment supports vLLM, SGLang, and TokenSpeed. The model is available via API on platform.kimi.ai.

What commenters are saying

Commenters note the model's size makes local inference challenging, with one calculating it needs roughly 1.5TB VRAM, though MoE streaming from SSD on 192GB systems is feasible. Several point out the licensing only affects model-as-a-service providers above $20M monthly revenue, leaving internal use free. Some question enforceability of the license in US courts, while others note Chinese companies have successfully sued in American courts. A few express hope that power infrastructure bottlenecks could eventually flood the market with datacenter GPUs.