Moonshot releases Kimi K3.1 Pro, a 3T open-weight model that nears Claude Opus 5.5 on coding

What Moonshot released

Moonshot AI on 23 October 2026 introduced Kimi K3.1 Pro, the first model in its K3.1 series. The company calls it its most capable 3T-class open-weight model so far, with a context window of one million tokens and gains in coding, 3D creation, financial research and other agent work.

Alongside it comes Kimi K3.1 Pro Highspeed, which Moonshot says runs two to three times faster for more responsive use. Both are available now in Kimi's apps and API. The full model weights are not out yet: Moonshot says they will be released by 4 November 2026.

The headline claim is that an open-weight model now sits within touching distance of the best proprietary systems on demanding agent tests. On Terminal-Bench 4.0, a test of coding agents working in resource-limited terminals, Moonshot reports 65.7% for K3.1 Pro, ahead of OpenAI's GPT-6 Astra at 58.2% and just behind Anthropic's Claude Opus 5.5 at 66.4%.

How it was built

K3.1 Pro is not a new base model. Moonshot describes it as a post-trained version of Kimi K3 that received substantially more post-training compute, mostly reinforcement learning across hundreds of downstream domains at once.

To make that work, the company says it optimised every layer of its stack: the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and a shared platform for checking the quality of training environments. The goal, in its words, was that each extra unit of RL compute delivers stable, predictable returns, and Moonshot says the result is consistent gains in every domain.

That matters for people already running K3. Since the architecture is unchanged, serving work such as vLLM's agent-assisted kernels that sped up Kimi K3 on NVIDIA GB300 should carry over once the K3.1 weights appear, although Moonshot has not said so explicitly.

Coding and long-horizon agents

Moonshot pitches K3.1 Pro as covering the full range of coding, from quick fixes in real codebases to building whole projects. On its internal Kimi Code Bench 3.0 Long-horizon, which tests multi-step execution, tool use and end-to-end problem solving, it scores 61.2%, up from 55.95% for K3. The company also says the model works across game engines such as Unity, Unreal Engine and Godot, and produces more polished front-end pages and Three.js scenes.

The most detailed example is a long internal test. Working on top of human-atlas, an open-source anatomical model with 296 skeletal components and 402 muscle components, K3.1 Pro built a human walking simulator with training and evaluation pipelines. Across 167 missions and 420 independent reviews, Moonshot says, it took the simulated person from collapsing within seconds to walking with alternating strides for 11.2 seconds.

The company highlights one moment in particular. When reward tweaks stopped improving arm motion, the model did not keep adjusting rewards; it probed the simulator's dynamics, found and fixed 36 errors in how shoulder-muscle moment arms were calculated, and retrained the controller. It is a single company-run example, but it is the kind of debugging behaviour long-running coding agents are judged on.

3D modelling and design

Moonshot puts unusual weight on 3D work. On BenchCAD, which asks models to rebuild 3D objects from multi-view renders by writing CAD code, K3.1 Pro scores 92.7% with tools, close to Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%. The company says it is especially strong at parametric modelling, from mechanical parts and architecture to configurable products and furniture.

It says the model can write and run scripts for Blender and CAD software directly, without connector plug-ins, to build and adjust geometry, materials and scene layout. Moonshot also launched 3D Make on kimi.com and the Kimi app, where users can preview and edit 3D models, download them for 3D printing, or export them as videos and Live Photos.

Finance and fewer hallucinations

For finance, Moonshot claims a state-of-the-art 64.4 on OfficeQA Pro v2, which tests analysis of real fiscal documents through retrieval, cross-checking and multi-step arithmetic. Its own finance tests show K3.1 Pro ahead of K3 by 7.2 points on financial modelling and 11.6 points on auditable research, and at about 4.2 times K3's score on a simulated business-management task.

Because finance punishes made-up facts, Moonshot also targeted factual accuracy. On Kimi Hallu Bench Hard, an internal stress test built from prompts where models tend to slip, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. The share of answers with at least one factual error dropped to 42.4% of K3's rate. Moonshot is careful to say these stress-test figures are not everyday error rates.

Holding back cyber capability on purpose

The most unusual part of the post is about cyber capability. Moonshot says vulnerability-research skill kept rising as post-training scaled, and traced it to cyber content inside the general mid-training data and cyber tasks in the RL environment mix. For the open-weight release, it removed cyber-related data from both stages while keeping general coding tasks.

On xdev-bench-hard, its in-house set of exploit tasks that recreate recent bugs in the Linux kernel, Node.js, JavaScriptCore, Exim and VirGL with mitigations on, a checkpoint trained without these filters solved 90.5% of problems at three attempts. The released K3.1 Pro solved 81.0%. So the filtering cut capability, but the model is still much stronger than K3, which Moonshot attributes to better general coding and debugging.

Moonshot says K3.1 Pro's offensive cyber ability stays behind proprietary frontier models that have been public for at least three months. It sets out two principles: do not release open weights that push the absolute frontier of cyber capability, and raise the level available in open weights gradually so defenders can audit code on infrastructure they control. Outside work such as Goodfire's cyber monitors built on Kimi K3 shows how researchers are already probing what these open models can do.

Where to get it

K3.1 Pro is rolling out on kimi.com and the Kimi mobile apps for iOS, Android and HarmonyOS, in the Kimi Work desktop app from version 3.1.0 on Windows and Apple silicon Macs, and in Kimi Code, where users switch with the /model command. Developers can call kimi-k3.1-pro or kimi-k3.1-pro-highspeed through the Kimi API, where zero data retention is available.

Kimi Enterprise administrators have to switch the model on for their organisation, because access is off by default at launch. Moonshot did not list API prices for the new models in the launch post.

Caveats on the numbers

Almost every number here comes from Moonshot. Several benchmarks, including Kimi Code Bench, Kimi Hallu Bench Hard, xdev-bench-hard and the finance suites, are internal and not public. Kimi's results use maximum reasoning effort, and many competitor scores are copied from other companies' reports or leaderboards rather than rerun under identical conditions; Moonshot itself notes proprietary model tests may involve fallback behaviour.

Some press coverage reports an early fourth place on Artificial Analysis's intelligence index, but AI Lookout has not confirmed that ranking independently. The real test will come after 4 November, when anyone can download the weights and run their own evaluations.

Why it matters

Open-weight models from Chinese labs keep closing the gap with US frontier systems. K3.1 Pro arrives a week after DeepSeek V4.1-Flash-1017 and claims scores within a point or two of Claude Opus 5.5 on coding and CAD tests, at a time when OpenAI has accused Moonshot and others of distillation of its models.

The cyber section may matter as much as the benchmarks. Moonshot is the first major open-weight lab to publish a before-and-after measurement of deliberately removing cyber training data, and to state a policy of staying behind the proprietary frontier. Whether outside red-teamers agree once the weights ship will shape how regulators view open releases of this size.