Moonshot AI launched Kimi K3.1 Pro on 23 October. The tech blog calls it “the first model in our new K3.1 series and our most capable 3T-class open-weight model to date,” with a 1-million-token context window. A second model, Kimi K3.1 Pro Highspeed, runs at 2 to 3 times the speed of K3.1 Pro, per Moonshot.
The word “open-weight” needs a date next to it. The blog says: “The full model weights will be released by Nov 4.” When we checked the moonshotai page on Hugging Face on 23 October, Kimi-K3 was still the newest model there.
Where you can use it today
Per the blog, both models roll out today on Kimi.ai and the mobile apps, the Kimi Work desktop app (version 3.1.0 or later), Kimi Code (pick it with /model) and the Kimi API, as kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Zero Data Retention is available to API customers. For Kimi Enterprise, access is off by default until an administrator turns it on.
Moonshot’s launch post on X gives the size as 2.8T total and 104B active parameters in a mixture-of-experts design, and says the model is multimodal. The API pricing page lists K3.1 Pro under the K3 series, where every model gets a 1,048,576-token context window. The per-token prices did not show in the page text we could read, so check them there before you budget.
What changed from K3
The blog says K3.1 Pro is a post-trained version of Kimi K3 “with substantially larger post-training compute,” using large-scale reinforcement learning across hundreds of domains. It pitches three areas:
- Coding. Full-project builds, including games in Unity, Unreal Engine and Godot, plus Three.js scenes and front-end pages.
- 3D. Parametric models in Blender and CAD, written as scripts the model runs itself. A “3D Make” feature is now live on kimi.ai and the app.
- Finance. Spreadsheet models, investment analysis and source-grounded research.
Moonshot’s numbers
All of these are Moonshot’s own runs or its citations of other labs’ reports:
- Terminal-Bench 4.0: 65.7%, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. Moonshot ran K3.1 Pro with the Claude Code harness.
- Kimi Code Bench 3.0 Long-horizon (internal): 61.2%, up from 55.95% for K3.
- BenchCAD with tools: 92.7%, against 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra.
- OfficeQA Pro v2: 64.4, which Moonshot calls state of the art.
- Kimi Hallu Bench Hard (internal): the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. Moonshot adds that the set is a deliberate stress test and “the reported rates do not represent overall error rates in daily use.”
The blog also describes one internal long-horizon run: building a human walking simulator, where the model found and fixed 36 errors in shoulder-muscle calculations. That is a single example from Moonshot, not a benchmark.
The cyber section
Moonshot says it removed cyber-related data from mid-training and from the reinforcement-learning mix for the open-weight release. On its in-house exploit benchmark, xdev-bench-hard, a checkpoint without those cuts solved 90.5% of tasks at pass@3, and the shipped K3.1 Pro solved 81.0%. Moonshot says the model’s cyber capability “remains behind that of frontier proprietary models that have been publicly available for at least three months,” and that it aims not to release open weights that “extend the absolute frontier of cyber capability.” These are Moonshot’s own measurements, and nobody outside the company has checked them yet.
If you are on K2
The pricing page says K2 series models will not be available to newly registered users from 31 October 2026, and that Moonshot plans to retire all K2 models by the end of December 2026. Existing users keep access until then. If you pin a K2 model in an app, plan the move now.
Update, 26 October 2026: Moonshot apologises for subscription limits and resets quotas
Moonshot apologised on X on 25 October “for the rate limits and unnecessary quota consumption on the K3.1 Pro subscription plan,” adding: “These were our issues, and we should have fixed them sooner.”
Every subscriber on every plan gets a one-time usage reset that restores 25% of the monthly quota, which Moonshot says equals one week of usage. Anyone who had used less than 25% goes back to zero. Moonshot also says it has adjusted the plan and is cutting unnecessary subagent calls and output length so quota lasts longer. Reuters notes Moonshot also had compute capacity problems after the K3 launch. The open-weights date in this post is unchanged.
Update, 4 November 2026: the weights are public, under a license with a Model-as-a-Service clause
Moonshot met its date. The Kimi-K3.1-Pro repository on Hugging Face was public and ungated when we checked on the evening of 4 November, with the weights in 96 safetensors files. The model card describes it as “an open-weight, natively multimodal agentic model” with 2.8T total and 104B activated parameters, a 1,048,576-token context, text and image input, and a 401M-parameter MoonViT-V2 vision encoder. Its quantization row reads “MXFP4 weights / MXFP8 activations (quantization-aware training),” and the card is tagged compressed-tensors. The repository includes its own configuration and processor code files, and Hugging Face tags it custom_code. The card recommends vLLM, SGLang and TokenSpeed for serving.
The license is not plain MIT. The Kimi K3.1-Pro License grants broad rights to use, modify, fine-tune and sell, “subject to the following conditions.” The key one is Section 2. It defines “Model as a Service” as “giving a third party access to language model inference or fine-tuning (e.g., via API) in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data,” excluding end-user products where the model is “solely embedded within specific features or harnesses” and “mere relaying of requests to models hosted by others.” Then: “If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or the equivalent in other currencies) in total 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.”
Section 3 adds an attribution rule: commercial products or services with “more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue” must display “Kimi K3.1-Pro” prominently in the user interface. Section 4 exempts internal use, defined as use that does not make the software, its outputs or its capabilities available to third parties, and use through Moonshot’s official products or certified inference partners.
If you only run the model inside your own company, Sections 2 and 3 do not apply. If you resell inference and your group’s revenue passes US$20 million over any 12 months, you need a separate deal with Moonshot first, and as written the threshold counts the aggregate revenue of the licensee and its affiliates, not just the inference business. The benchmark figures above are still Moonshot’s own, and now anyone with the hardware can check them.

The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment