Moonshot AI published the open weights of Kimi K3.1 Pro, a 2.8-trillion-parameter model

Moonshot AI’s Hugging Face repository moonshotai/Kimi-K3.1-Pro holds the full weights of Kimi K3.1 Pro, the model Moonshot launched by API in late October. The repository’s initial commit is dated 2026-11-02 and its latest update 2026-11-05, the day TechNode reported the weights as released. It follows Kimi K3, introduced in July with weights promised for July 27.

According to the model card, K3.1 Pro is a mixture-of-experts model with 2.8 trillion total parameters and 104 billion activated, 93 layers, and 896 experts of which 16 are selected per token plus 2 shared. It has a 1,048,576-token context window and takes text and image input through a 401M-parameter MoonViT-V2 vision encoder. It was trained with quantization-aware training, shipping MXFP4 weights with MXFP8 activations, and the card recommends vLLM, SGLang or TokenSpeed for serving.

Moonshot’s reported scores at max reasoning effort include 94.2 on GPQA-Diamond, 92.1 on BrowseComp, 79.6 on OSWorld V2.1, 75.5 on DeepSWE v1.1, 70.0 on HLE-Full with tools and 65.7 on Terminal-Bench 4.0, several of them run with the Claude Code harness. The weights ship under a model-specific Kimi K3.1 Pro License rather than a standard open-source license.

Why it matters: a model in the same class as the largest frontier systems is again downloadable within weeks of its API launch, keeping the open-weight frontier only a short step behind the closed one.

What it does not show: Hugging Face shows commit dates, not the moment the repository became public, so the exact release day is uncertain; the benchmarks are Moonshot’s own; and the custom license, not the parameter count, decides what businesses may actually do with it.