Kimi K3.1 Pro open weights are out on Hugging Face

Moonshot AI released the full open weights of Kimi K3.1 Pro on Hugging Face today (Nov 4, 2026), matching the date it promised in the Oct 23 launch blog (opens in a new tab). @Kimi_Moonshot posted the drop with a link to the repo and the line "We'll keep advancing open intelligence." Our earlier story covered the app/API launch when the weights were still due; this one is about the download itself.

What shipped. The Hugging Face repo moonshotai/Kimi-K3.1-Pro (opens in a new tab) includes a model card, LICENSE, config, custom code, and 96 safetensors shards (model-00001-of-000096 …). Moonshot recommends serving with vLLM, SGLang, or TokenSpeed. The hosted API remains on platform.kimi.ai (opens in a new tab) as kimi-k3.1-pro; Together AI (opens in a new tab) also lists the model for serverless/dedicated inference.

Specs (from the model card). Mixture-of-experts: 2.8T total / 104B active parameters, 1,048,576-token context, multimodal text + image input, MoonViT-V2 vision encoder (401M params), and MXFP4 weights / MXFP8 activations via quantization-aware training. Moonshot calls it its most capable 3T-class open-weight model to date, post-trained from Kimi K3 with large-scale joint RL.

Who it's for. Labs and self-hosters who already run huge MoE stacks and want a frontier-class open checkpoint for long-context coding, agent work, and multimodal jobs — not anyone expecting to load this on a laptop. For most builders, the Kimi API (or a host like Together) is still the practical path.

License caveat. The file is titled Kimi K3.1-Pro License (HF tag license:other). It reads MIT-like, but: (2) if you operate a Model as a Service business and you and your affiliates exceed $20M revenue over any consecutive 12 months, you need a separate Moonshot agreement before commercial use; (3) commercial products with >100M MAU or >$20M monthly revenue must display "Kimi K3.1-Pro" prominently; (4) internal use and use via Moonshot official products or certified partners are exempt from (2) and (3). Read the LICENSE before you ship.

Benchmarks. The HF card lists many scores (e.g. Terminal-Bench 4.0 at 65.7%). Those are Moonshot's own numbers unless a footnote cites another source — treat them as vendor results, not independent audits.

Honest caveats. This is not a laptop model. Serving a same-size Kimi K3 class checkpoint has historically meant multi-GPU / multi-node datacenter gear. MODEL has not tested Kimi K3.1 Pro or these weights.

Disclosure: Not sponsored. No affiliate links. Not investment advice.

MODEL Editors