Moonshot Releases Open Weights for Kimi K3.1 Pro

At a glance

The weight-release announcement is November 4, 2026, from the @Kimi_Moonshot post at 16:00 UTC. The prior commitment was full weights by November 4, 2026, from the Kimi tech blog. The repository is moonshotai/Kimi-K3.1-Pro on Hugging Face, public and not gated, from the model card. Scale is 2.8T total and 104B active parameters in a mixture of experts, with a 1,048,576-token context, from the model card. Modalities are text and image in, with a MoonViT-V2 vision encoder of 401 million parameters, from the model card.

Quantization is MXFP4 weights and MXFP8 activations, with quantization-aware training, from the model card. The license is the Kimi K3.1-Pro License: a Model-as-a-Service business above $20 million over 12 months needs a separate deal, and large commercial products must show attribution, from the LICENSE file. Recommended serving stacks are vLLM, SGLang, and TokenSpeed, and the API model id remains kimi-k3.1-pro, from the model card and the blog.

What shipped

The news is the weight drop, not a new model launch. Moonshot’s earlier Introducing Kimi K3.1 Pro tech blog presented K3.1 Pro as the first model in the K3.1 series and “most capable 3T-class open-weight model to date,” with apps and the API already live and full weights promised by November 4. my ai daily covered that promise in “Kimi K3.1 Pro arrives with 1M context; weights promised by November 4.” The November 4 X post from @Kimi_Moonshot points directly at the Hugging Face repo and restates the open-intelligence line from the blog’s framing.

The model card describes an open-weight, natively multimodal agentic mixture-of-experts model post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains. Versus K3, it replaces full-attention Gated MLA layers with Gated DeepSeek Sparse Attention (Gated DSA) for cheaper long context: 69 KDA layers plus 24 Gated DSA layers (93 total), 896 experts with 16 selected per token plus 2 shared experts.

Hugging Face lists the repo as public and ungated. The card recommends running local inference with vLLM, SGLang, or TokenSpeed. Hosted API access continues via kimi-k3.1-pro on Moonshot’s platform. The launch blog also shipped Kimi K3.1 Pro Highspeed (about 2–3 times the speed of Pro) in products; this weight release’s public repository name is Kimi-K3.1-Pro. Moonshot has not published a separate Highspeed weight repo in the sources checked for this article.

License boundaries

Weights ship under the Kimi K3.1-Pro License, copyright 2026 Moonshot AI. Broad use, modification, fine-tuning, redistribution, and commercial use are allowed, with two commercial thresholds: operators of a Model as a Service business whose aggregate revenue exceeds $20 million over any consecutive 12 months need a separate Moonshot agreement before commercial use of the Software or derivatives; and commercial products or services with more than 100 million monthly active users or more than $20 million in monthly revenue must display “Kimi K3.1-Pro” prominently. Internal use and use through Moonshot’s official products or certified inference partners are carved out of those two conditions.

Capability claims to keep attributed

Benchmark tables on the model card and tech blog are Moonshot’s (and, where footnoted, third-party leaderboards Moonshot cites). Headline self-reported figures include 65.7% on Terminal-Bench 4.0 (ahead of GPT-6 Astra’s 58.2%, behind Claude Opus 5.5’s 66.4% on Moonshot’s comparison), 61.2% on internal Kimi Code Bench 3.0 Long-horizon (versus K3 at 55.95%), 92.7% BenchCAD with tools, and 64.4 on OfficeQA Pro v2. On internal Kimi Hallu Bench Hard, Moonshot reports claim-level hallucination down from K3’s 23.4% to 8.3%, with an explicit stress-test caveat that rates are not overall daily-use error rates.

On cybersecurity, the launch blog says Moonshot removed cyber-related data from mid-training and reinforcement learning for the open-weight release. On internal xdev-bench-hard, it reports 81.0% pass@3 for the final model versus 90.5% for a checkpoint without those removals, and says offensive cyber capability remains behind frontier proprietary models that have been public for at least three months. Treat those as first-party risk-reduction claims, not independent audits.

Bottom line

Kimi K3.1 Pro’s downloadable weights are now public on Hugging Face as Moonshot promised, under a permissive but thresholded custom license, with the same 2.8T / 104B / 1M-context multimodal card Moonshot published at launch. The actionable change for builders is local and self-hosted access; product and API availability were already live before the weight drop.