Moonshot AI Publishes Kimi K3.1 Pro Weights Under a License Requiring a Separate Agreement for Model-as-a-Service Businesses Above $20 Million

Overview

Moonshot AI has published the weights of Kimi K3.1 Pro, according to TechNode, which describes the model as the company’s flagship multimodal model and says the weights are on Hugging Face. The Hugging Face model card says the model is released under the Kimi K3.1 Pro License. The model was already available through Moonshot’s apps and API before the weights appeared, and an independent index operator labels it proprietary. This article covers the release, the license terms as written in the repository, and a gap between two published benchmark scores.

What We Know

Release timeline. Moonshot’s technical blog post calls Kimi K3.1 Pro “the first model in our new K3.1 series” and says it and a faster Kimi K3.1 Pro Highspeed variant were available on Kimi.com, Kimi Work, Kimi Code and the Kimi API, adding: “The full model weights will be released by Nov 4.” The Decoder reported the launch on October 23, 2026. TechNode, dated November 5, 2026, says the weights release follows the model’s earlier availability through Kimi’s applications and API, and that Huawei has announced inference support through vLLM Ascend and SGLang, covering its Atlas A3 and Ascend 950 systems.

Specifications. The model card lists a Mixture-of-Experts architecture with 2.8T total parameters and 104B activated parameters, a context length of 1048576, MXFP4 weights with MXFP8 activations, and text and image modality. It recommends running the model on vLLM, SGLang or TokenSpeed. For API use it names the model kimi-k3.1-pro on platform.kimi.ai and says the thinking effort is set through a top-level reasoning_effort request field that supports “low”, “high” and “max”, with “max” as the default.

API pricing. As The Decoder reported on October 23, 2026, Moonshot’s token prices fell by 33 percent: input costs $2 per million tokens (down from $3), output $10 (down from $15), and cache hits $0.20 (down from $0.30). The Herald did not re-read Moonshot’s pricing page for this article, so current prices may differ. Artificial Analysis lists $2.00 per 1M input tokens and $10.00 per 1M output tokens as read on November 5, 2026.

The License

The license file in the repository is titled Kimi K3.1-Pro License and is copyrighted to Moonshot AI. It grants permission to deal in the Software, which it defines to include the model weights, parameters, configuration files, and inference and training code, “without restriction”, subject to numbered conditions. As written in the file:

• Model as a Service threshold (section 2). The license defines Model as a Service as giving a third party access to language model inference or fine-tuning, for example via an API, in a way that lets that party exercise meaningful control over inputs, parameters or training data. It excludes end-user products with model capabilities solely embedded within specific features or harnesses, and mere relaying of requests to models hosted by others. If a licensee or any of its affiliates operates a Model as a Service business and their aggregate revenue exceeds 20 million US dollars over any consecutive 12 months, the licensee must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose.
• Display requirement (section 3). For commercial products or services with more than 100 million monthly active users, or more than 20 million US dollars in monthly revenue, the license says the name Kimi K3.1-Pro “must be prominently displayed on the user interface”.
• Exemptions (section 4). Sections 2 and 3 do not apply to internal use, defined as any use that does not make the Software, its outputs or its underlying capabilities available to third parties, or to use of the Software accessed through Moonshot AI’s official products or certified inference partners.

The Herald previously reported on another recent release with a Model-as-a-Service condition, in Mistral Releases Large 4 Open Weights Under a Modified MIT License That Requires a Commercial Agreement for Model-as-a-Service Businesses. It earlier covered the predecessor in Moonshot AI Releases Kimi K3, a 2.8-Trillion-Parameter Open-Weight Model That Rattles Chinese Tech and Chip Stocks.

Cyber-Capability Controls

Moonshot’s blog post says it traced improving vulnerability-research performance to cyber-related content in the general mid-training corpus and in the reinforcement-learning environment mix, and that for the open-weight release it “identified and removed cyber-related data from both training stages”. On its internal xdev-bench-hard, Moonshot reports that the checkpoint without these measures solves 90.5% of problems at pass@3 and the final model reaches 81.0%. The post also says the model’s cyber capability “remains behind that of frontier proprietary models that have been publicly available for at least three months.” The Decoder repeats the 90.5 and 81.0 percent figures and says DeepSeek-V4.1-Flash reaches 85.7 percent on the same test. These are Moonshot’s own benchmark figures, and the Herald found no independent replication of them in the sources reviewed.

Two Benchmark Scores

The Decoder reported that Kimi K3.1 Pro scores 55 points on the Artificial Analysis Intelligence Index, an 11-point jump over Kimi K3, placing it fourth behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56), according to Artificial Analysis. The Decoder wrote that if Moonshot released the weights as planned by November 4, 2026, the model “would become the highest-scoring open-weights model available.”

The Artificial Analysis model page, as read on November 5, 2026, shows a different figure: it says Kimi K3.1 Pro scores 46 on the Intelligence Index, above a median of 37 for comparable models, and cites Intelligence Index v5.0.1 with nine named evaluations. The same page is headed “Proprietary model” and “Released October 2026”. Neither The Decoder nor the model page explains the difference between 55 and 46, and the Decoder article does not state an index version, so the Herald cannot say what accounts for it. The page’s own filter definition says the open-weights label indicates “whether the model weights are available”, so the proprietary label may not yet reflect the Hugging Face release, but the page does not say so.

On Terminal-Bench 4.0, The Decoder reports that Moonshot’s 65.7 percent compares with 58.2 percent for GPT-6 Astra and 66.4 percent for Claude Opus 5.5. That is a vendor-reported figure.

What We Don’t Know

• Whether Artificial Analysis will relabel the model as open weights, and whether its score will change.
• Whether any independent party has reproduced Moonshot’s coding, finance or cyber figures. The sources reviewed did not report one.
• How the Model as a Service revenue threshold will be applied in practice. The license text defines terms, but the sources reviewed contain no commentary on enforcement.
• Whether Moonshot’s API prices have changed since the October 23 reporting.