Kimi K3.1 Pro: Chinese model jumps to No. 4, open weights to follow

What this is about

Chinese AI company Moonshot AI has released Kimi K3.1 Pro, the first model in a new K3.1 series. It is a retrained version of Kimi K3, and Moonshot says it put significantly more compute into post-training. Post-training is the phase after the main base training, in which the model is refined, for example with reinforcement learning (learning by being rewarded for solving tasks).

On the Artificial Analysis Intelligence Index, an independent overall comparison of AI models, K3.1 Pro scores 55. That is 11 points more than its predecessor and good enough for 4th place. Only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56) rank higher.

There is also a K3.1 Pro Highspeed variant, which Moonshot says is two to three times faster with the same capabilities. Both models are available now through Kimi.ai and the Kimi apps (iOS, Android, HarmonyOS), the Kimi Work desktop app (version 3.1.0 and later, for Windows and Apple-silicon Macs), Kimi Code and the Kimi API. In the API they are called kimi-k3.1-pro and kimi-k3.1-pro-highspeed. API customers get Zero Data Retention. In Kimi Enterprise, access is disabled by default at launch and must be enabled by administrators.

Size, pricing and availability

K3.1 Pro keeps K3's dimensions: 2.8 trillion parameters in total, 104 billion of them active. Parameters are, roughly speaking, the dials a model tunes during training. "Active" means that only a portion of them is used for each request, as is common in mixture-of-experts architectures. The context window, i.e. the amount of text the model can process at once, is one million tokens. The model handles text and images, and according to the Kimi Code documentation it also accepts video.

Prices drop by 33 percent:

K3 previously cost $3 for input, $15 for output and $0.30 for cache hits. Running the Intelligence Index costs $1.98 with K3.1 Pro, almost the same as with K3 ($2.00), because the model generates 14 percent more output tokens. According to Artificial Analysis, it sits at the front of the Pareto frontier: no model with a comparable parameter count scores higher on the index.

From the Kimi Code documentation: K3.1 models support thinking mode only. Requests without thinking are rejected with an HTTP 400 error. The reasoning_effort parameter accepts low, high and max, with max as the default. Depending on your plan, the models are available only from Plus or Pro upward, and the million-token context is reserved for Pro and above.

Benchmarks: biggest jumps on agentic tasks

Agentic tasks, where the model works independently with tools over multiple steps, show the biggest gains. On GDPval-AA, K3.1 Pro reaches 1,846 Elo according to Artificial Analysis, 312 more than K3 and just behind Claude Opus 5.5 (1,866) and Muse Wonder Full (1,857). On Terminal-Bench 4.0, Moonshot reports 65.7 percent, while K3 managed only 13 percent.

Selection from Moonshot's table (all at reasoning effort "max", vendor-reported):

On Kimi Code Bench 3.0 Long-horizon, an internal test, the model scores 61.2 percent versus 55.95 percent for K3. On internal finance benchmarks, it improves by 7.2 points on Financial Modeling and 11.6 points on Financial Research.

The hallucination rate (fabricated claims) also falls: on the internal Kimi Hallu Bench Hard, it dropped from 23.4 to 8.3 percent at the claim level. Moonshot stresses that this is a stress test, not a general error rate. GPT-6 Astra scores 5.5 percent there.

As an example of long-running tasks, Moonshot cites an internal test: over 167 missions and 420 independent reviews, the model built a human gait simulator based on an anatomical model with 296 skeletal and 402 muscle components. Along the way, it found and fixed 36 errors in shoulder muscle calculations.

3D modeling and coding

According to Moonshot, K3.1 Pro works directly with game engines such as Unity, Unreal Engine and Godot, from quick fixes to building a complete project. On BenchCAD, which tests 3D reconstruction from views via CAD code, it scores 92.7 percent with tools and 72.8 percent without (Claude Opus 5.5: 73.0 percent). It is said to be especially strong at parametric modeling, such as mechanical parts, architectural models and configurable products. It can write and run scripts for Blender and CAD software without MCP connectors. Also new is "3D Make" on Kimi.com and in the app, where you can view and edit 3D models and download them for 3D printing.

Cybersecurity: deliberately limited capabilities

Moonshot devotes a separate section to cybersecurity. The company observed that vulnerability-hunting ability rose steadily with more post-training, because cyber content was part of the general mid-training dataset and corresponding tasks were in the reinforcement learning mix. For the open-weights version, Moonshot removed this data from both training phases but kept general coding tasks.

On the internal xdev-bench-hard, which recreates recently disclosed vulnerabilities in the Linux kernel, Node.js and JavaScriptCore, among others, the unfiltered version solves 90.5 percent of tasks (pass@3), while the final K3.1 Pro solves 81.0 percent. DeepSeek-V4.1-Flash scores 85.7 percent and Kimi K3 33.3 percent. Moonshot says its cyber capabilities remain behind proprietary frontier models that have been available for at least three months.

Moonshot sets out two principles: open weights should not push the absolute frontier of offensive cyber capabilities, but available capabilities should rise gradually, because open models also serve defense, for example when auditing legacy systems.

Our take

If Moonshot releases the weights by November 4 as announced, K3.1 Pro would be the highest-rated open-weights model, according to Artificial Analysis. The gap to proprietary models would shrink from 12 to 3 points. For now, though, the weights are not out. Unite.AI calls the model "open-weight", but the release is still pending. A model of this size can only be run locally on server hardware anyway.

A few caveats: most benchmark figures come from Moonshot itself, and internal tests such as Hallu Bench or xdev-bench-hard have not been independently verified. Competitors lead on GPQA-Diamond and ExploitBench, and GPT-6 Astra also leads on hallucinations. Filtering training data against cyber capabilities is an interesting approach, but so far only Moonshot's own test supports its effectiveness.