Image: Moonshot AI

Moonshot AI has launched Kimi K3.1 Pro, a new version of its largest model. The Chinese company calls it its most capable open-weight model so far. Open weight means anyone will be able to download the model and run it on their own hardware. Moonshot says it will publish those files by November 4, 2026.

For now, the model is live in the Kimi apps, the Kimi Work desktop app, Kimi Code and the Kimi API. A faster version called K3.1 Pro Highspeed is also out. Moonshot says it runs two to three times faster.

Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.

🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.

Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.

Open weights by Nov 4, 2026. pic.twitter.com/klydcszv02

— Kimi.ai (@Kimi_Moonshot) October 23, 2026

According to Moonshot's tech blog, K3.1 Pro has 2.8 trillion parameters in total. Parameters are the internal settings a model learns during training. Only 104 billion are active for any single answer, because the model uses a mixture-of-experts design. That splits the model into specialist parts and wakes up only the ones a task needs. It can also take in images and handle about one million tokens of context, roughly several long books at once.

Moonshot says the model is a heavily retrained version of Kimi K3. The company says it improved at coding whole projects, including games in Unity, Unreal Engine and Godot. It also claims gains in 3D modeling with Blender and CAD software, and in finance work such as building spreadsheet models.

On Terminal-Bench 4.0, a test of coding agents working in a command line, Moonshot reports a score of 65.7%. It says that beats OpenAI's GPT-6 Astra at 58.2% and trails Anthropic's Claude Opus 5.5 at 66.4%. These are Moonshot's own numbers. Several other results come from Moonshot's internal tests, which outsiders cannot easily check.

One outside group has weighed in. Artificial Analysis, which runs its own benchmarks, placed K3.1 Pro fourth on its Intelligence Index with a score of 55. Only three models scored higher. The group says K3.1 Pro will be the top open-weight model on its index once the weights are out. It also measured 67% on Terminal-Bench 4.0.

Moonshot also says the model makes things up less often. On its own stress test, the share of false claims fell from 23.4% for K3 to 8.3%. The company notes that the test is built to trip models up, so the figures do not reflect everyday error rates.

Image: Moonshot AI

The most unusual part of the launch is about hacking skills. Moonshot says it removed cyberattack material from training before the open release. Its goal, it says, is to avoid giving away a free model that pushes the limits of what attackers can do. On its in-house exploit test, a version trained without those cuts solved 90.5% of tasks. The released model solved 81%. The Next Web points out that this is still far above K3's 33.3%.

Image: Moonshot AI

Moonshot says K3.1 Pro still trails the best closed models at cyber tasks. It argues that open access helps defenders audit their own code. That claim, like the benchmark results, has not been independently tested.

Tips if you want to try it

The model is available today on kimi.ai and in the Kimi mobile apps. Developers can pick kimi-k3.1-pro or kimi-k3.1-pro-highspeed on the Kimi API platform, which needs a top-up of at least $1. Artificial Analysis lists the price at $2 per million input tokens and $10 per million output tokens. Business admins need to switch the model on, because access is off by default. Anyone planning to run it locally should wait for the weights and check the license terms.