Moonshot releases Kimi K3.1 Pro: a 3T open-weight model with a 1M-token context window

Moonshot AI has unveiled Kimi K3.1 Pro, the first model in the K3.1 series and, according to the company, the most powerful open-weight model in the 3T-parameter class. The context window is 1M tokens. On Terminal-Bench 4.0 it scores 65.7%, more than GPT-6 Astra (58.2%) and only slightly less than Claude Opus 5.5 (66.4%). K3.1 Pro Highspeed, which runs 2–3 times faster, launched with it. Both versions are already available in Kimi.ai, Kimi Work, Kimi Code and through the API, and Moonshot promises to publish the full weights by November 4, 2026.

Post-training as the main bet

K3.1 Pro is a further-trained version of Kimi K3, but with a noticeably larger compute budget for post-training. To run large-scale joint RL across hundreds of applied domains at once, the team reworked the entire stack: the RL framework, long-context inference, training efficiency, sandboxes, rollout scheduling and a shared platform for verifying the quality of environments. The goal is for every additional chunk of RL compute to deliver a steady, predictable gain.

On the internal Kimi Code Bench 3.0 Long-horizon, which tests multi-step task execution and tool use, the model scored 61.2% versus 55.95% for K3. Moonshot's most vivid example: using the human-atlas anatomical model (296 bone elements and 402 muscle elements), K3.1 Pro built a human walking simulator with training and evaluation pipelines. Over 167 missions and 420 independent reviews, the virtual human learned to walk for 11.2 seconds with alternating steps, instead of falling after a few seconds. When reward tuning failed to improve arm movement, the model did not keep tweaking it. It ran open-loop dynamics probes, found and fixed 36 errors in the moment-arm calculations for the shoulder muscles, and retrained the controller.

3D, CAD and finance

The second focus is spatial reasoning. On BenchCAD, where the task is to reconstruct a 3D object from multi-view renders by generating CAD code, K3.1 Pro with tools scores 92.7%. Claude Opus 5.5 gets 96.2% and GPT-6 Astra gets 95.9%. The model can work with Blender and CAD scripts without MCP connectors, and in Unity, Unreal Engine and Godot it builds games together with web search and 3D generation. For users, Moonshot launched 3D Make on kimi.ai and in the app: preview, editing, export for 3D printing, or export to video and Live Photos.

In finance, Moonshot claims a record on OfficeQA Pro v2 with 64.4. Internal metrics improve as follows:

• Financial Modeling: +7.2 points over K3
• Financial Research: +11.6 points
• Business Operation: roughly 4.2 times K3's result
• Kimi Hallu Bench Hard: the claim-level hallucination rate fell from 23.4% to 8.3%, and the share of answers with at least one factual error came to 42.4% of K3's rate

The company itself adds a candid caveat: the test is deliberately built from provocative examples and does not reflect the error rate in everyday use.

Cyber risks and the logic of open weights

A separate section of the release is devoted to cybersecurity. The team noticed that the ability to find vulnerabilities grows with post-training, because cyber content enters both the general mid-training corpus and the RL environments. For the open version, such data was removed at both stages, while ordinary programming tasks were kept. On xdev-bench-hard (exploiting real vulnerabilities in the Linux kernel, Node.js, JavaScriptCore, Exim and VirGL with protections enabled), a checkpoint without these measures solves 90.5% of tasks at pass@3, while the final K3.1 Pro solves 81.0%. In cyber capabilities, Moonshot says, it lags behind closed frontier models that have been available for at least three months.

The company's position is not to release open weights that extend the absolute frontier of offensive capabilities, but to raise the bar gradually. The argument for openness is code auditing, including of legacy systems that security specialists rarely review, carried out by many teams on their own infrastructure.

How to try it

• Kimi apps: iOS, Android and HarmonyOS, or kimi.ai
• Kimi Work: desktop version 3.1.0 or later for Windows and Mac on Apple silicon
• Kimi Code: select the model with the /model command
• API: model IDs kimi-k3.1-pro and kimi-k3.1-pro-highspeed; Zero Data Retention is available for customers
• Kimi Business: access to K3.1 Pro is off by default for administrators and must be enabled manually

When comparing with competitors, keep in mind that Moonshot took some of the figures from other labs' blogs and leaderboards and measured others itself: Terminal-Bench 4.0 was run with the Claude-Code harness, and PostTrainBench on H20 GPUs instead of H100. For teams that need agentic coding on their own infrastructure, the main reason to test comes after the weights are published: that is when they can run the model on their own repositories and compare it with closed APIs.