Moonshot AI has released the weights for Kimi K3.1 Pro on Hugging Face, describing it as "an open-weight, natively multimodal agentic model, and our most capable 3T-class model to date." The weights appeared this week and are distributed under the Kimi K3.1 Pro License.
According to the model card, K3.1 Pro is a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active per token. It has 93 layers, 896 experts with 16 selected per token plus 2 shared experts, and a 401-million-parameter MoonViT-V2 vision encoder. Weights ship in MXFP4 with MXFP8 activations after quantization-aware training, and the context window is 1,048,576 tokens.
Moonshot reports scores of 65.7 on Terminal-Bench 4.0, 75.5 on DeepSWE v1.1, 92.1 on BrowseComp, 94.2 on GPQA-Diamond, 79.6 on OSWorld V2.1, 51.1 on HLE-Full (70.0 with tools) and an Elo of 1846 on GDPval-AA v2.1. The card compares the model with Claude Fable 5.1, Claude Opus 5.5, GPT-6 Astra, DeepSeek-V4.1-Flash and GLM-5.3; it leads on some rows and trails on others, such as Terminal-Bench-Science, where GPT-6 Astra scores 68.1 against K3.1 Pro's 38.1.
The model is also served through Moonshot's API at platform.kimi.ai under the name kimi-k3.1-pro, with OpenAI- and Anthropic-compatible endpoints. Moonshot recommends vLLM, SGLang and TokenSpeed for self-hosting. Reasoning effort is set with a reasoning_effort field (low, high or max), and multi-turn or tool-calling conversations must pass back the full assistant message including reasoning_content.
At 2.8 trillion parameters, self-hosting requires multi-node hardware, but the release gives teams an open-weight option near the frontier for agentic and multimodal work, and the compatible API endpoints make it a drop-in candidate for existing agent stacks.