Moonshot publishes Kimi K3.1 Pro weights as promised: 2.8T parameters, million-token context, free to self-host

Chinese AI company Moonshot AI published the full weights of its flagship model Kimi K3.1 Pro on Hugging Face on November 4 (moonshotai/Kimi-K3.1-Pro), keeping the promise it made at the model's October 23 launch. Developers and research institutions can now deploy, study and adapt the model on their own infrastructure instead of relying only on Moonshot's hosted services. This site covered the launch on October 23; the weights release is the promised follow-through.

The model: 2.8 trillion parameters, million-token context

According to the model card on Hugging Face, Kimi K3.1 Pro is a mixture-of-experts (MoE) model with 2.8 trillion total parameters and 104 billion activated per token, across 93 layers and 896 experts, 16 of which are selected per token. Its context window is 1,048,576 tokens, about one million. The model is natively multimodal for text and images, uses the MoonViT-V2 vision encoder, and ships with MXFP4-quantised weights. Moonshot says it was post-trained from Kimi K3 with large-scale joint reinforcement learning, is the first model in the K3.1 series and the company's most capable 3-trillion-class model to date, aimed at coding, 3D creation, financial research and long-horizon agent work.

Company-reported scores: near the closed frontier on coding and agents

Moonshot's self-reported benchmarks put Kimi K3.1 Pro at 65.7% on Terminal-Bench 4.0, above GPT-6 Astra at 58.2% and Claude Fable 5.1 at 57.9%, and just below Claude Opus 5.5 at 66.4%; it scores 51.7 on SWE-Marathon, 64.4 on OfficeQA Pro v2 and 92.1 on BrowseComp. These are the company's own numbers, not independently verified, and real-world performance still awaits third-party reproduction. On the API side, the model is live at platform.kimi.ai as kimi-k3.1-pro with OpenAI- and Anthropic-compatible formats, priced at $2 per million input tokens and $10 per million output tokens; supported inference engines include vLLM, SGLang and TokenSpeed. A Highspeed variant generates output two to three times faster.

Licence and limits: cyber-offence data removed from the open release

The licence deserves a close read: the model ships under Moonshot's custom Kimi K3.1 Pro licence, broadly permissive but with conditions reserved for massive commercial services, a revenue-gated arrangement rather than unrestricted open source. Moonshot also removed cyber-offence-related data from both training stages of the open release, which cut its internal xdev-bench-hard score from 90.5% to 81.0%; the hosted model's behaviour may differ. In other words, the published weights are a safety-processed version, and researchers and companies should read the licence terms before adopting it.

(Sources: Moonshot AI Hugging Face model card; Dosa AI Tools briefing, November 5, 2026; ExplainX update, November 4, 2026; Latent Space AI News. Benchmark scores are company-reported. Cover photo is a file illustration of an AI chip.)