Moonshot AI published the full model weights of Kimi K3.1 Pro on Hugging Face on 4 November, opening up its most capable model so far. The announcement on X drew more than 10,000 views within a day, and the release thread on Hacker News climbed past 520 points.
Kimi K3.1 Pro is a natively multimodal mixture-of-experts model with 2.8 trillion total parameters and 104 billion active per token. It carries a 1 million token context window, a MoonViT-V2 vision encoder, and weights trained with MXFP4 quantization so they fit on far less hardware than the raw parameter count suggests. Moonshot reports reasoning and agentic scores close to the closed frontier: 65.7 on Terminal-Bench 4.0 against 57.9 for Claude Fable 5.1 and 66.4 for Claude Opus 5.5, 92.1 on BrowseComp ahead of GPT-6 Astra at 91.5, and 75.5 on DeepSWE v1.1.
For founders, the practical point is that frontier-class coding help no longer requires sending code to a closed lab. K3.1 Pro can be self-hosted or rented from providers such as Together AI, which matters if you handle client data, proprietary code, or regulated workloads. The trade-off is infrastructure: a 2.8T MoE model is not a laptop download, and long-context agent runs get expensive. A sensible next step is a side-by-side test of K3.1 Pro against whatever you run today on your own repository and prompts before switching anything in production.