Moonshot AI has published the full weights of Kimi K3.1 Pro on Hugging Face, delivering the open release it promised when the model launched on October 23 and giving researchers the largest open-weight model yet available on the platform. The repository, moonshotai/Kimi-K3.1-Pro, is public and not gated, so anyone can download the trained model and run it independently — provided they have the hardware to handle it.
The weights went live on November 4, the deadline Moonshot set at launch, and the model card describes Kimi K3.1 Pro as an open-weight, natively multimodal agentic model and Moonshot's most capable 3T-class model to date. It is the first release in the K3.1 series.
A 2.8-trillion-parameter model that runs lean
Kimi K3.1 Pro is a mixture-of-experts model with 2.8 trillion parameters in total, but only about 104 billion are active for each token. Each token is routed to 16 of the model's 896 experts, plus two shared experts, which is why a model of this size can answer at a cost closer to a far smaller one. It carries a context window of 1,048,576 tokens and handles text and images natively, with a 401-million-parameter MoonViT-V2 component handling vision.
The key architectural change from Kimi K3 is in the attention mechanism. In 24 of the model's 93 layers, full Gated MLA attention is replaced by sparse Gated DSA attention, which speeds up processing over very long contexts. The weights ship at MXFP4 precision after quantization-aware training, keeping the download to about 1.56 terabytes across 96 files — still enormous, but far smaller than the roughly 5.6 terabytes a 16-bit copy would require.
Near-flagship scores, with a safety edit first
Moonshot's company-reported scores put Kimi K3.1 Pro within a point of Anthropic's Claude Opus 5.5 on one of the toughest coding-agent tests: 65.7 percent on Terminal-Bench 4.0, against 66.4 percent for Opus 5.5 and 58.2 percent for OpenAI's GPT-6 Astra. The company also reports 92.1 on BrowseComp and 94.2 on GPQA-Diamond. Those figures are vendor-reported with no independent reproductions yet, so they should be read as Moonshot's claims rather than settled results.
The release also carries an unusual safety note. Moonshot said it traced steadily rising exploit skill in the model to cyber content buried in its general mid-training corpus and in its reinforcement-learning environments, and removed both before opening the weights. That filtering is now part of the release record, not an afterthought.
A data-centre download, not a laptop one
For all the openness, running Kimi K3.1 Pro locally is a data-centre job. Moonshot published deployment recipes for vLLM, SGLang and TokenSpeed, and the vLLM recipe runs the model across eight GPUs — two Nvidia GB300 trays, one B300 node or one AMD MI355X node. Most teams will still reach the model through Moonshot's API, where it costs $2 per million input tokens and $10 per million output tokens.
The licence deserves a careful read, too. The Kimi K3.1 Pro License follows the Kimi K3 pattern: free to use, change and sell, but a company running a model-as-a-service business with more than $20 million in revenue over 12 months must sign a separate agreement with Moonshot. It is a custom licence, not MIT or Apache.
Even with those caveats, the release marks a moment for open weights. At 2.8 trillion parameters, Kimi K3.1 Pro is the largest real model release on Hugging Face, ahead of Qwen3.8-2.4T and DeepSeek-V4-Pro at 1.6 trillion. The gap between the best open models and the best closed ones keeps narrowing — and now anyone with enough GPUs can inspect, fine-tune and run one of them.
