Moonshot AI's Kimi K3.1 Pro brings a 1 million-token context window

Moonshot AI, the company behind the Kimi chatbot, has released a new artificial intelligence model called Kimi K3.1 Pro. In its launch post, Kimi says the model can handle a context window of 1 million tokens, meaning it can take in very long documents or codebases in a single request.

Kimi calls K3.1 Pro "our most capable 3T-class open-weight model to date," according to its blog. The model is available now, but the downloadable weights are not out yet. Kimi says it will release the full weights by Nov. 4. Every benchmark score cited below was reported by Kimi, and none has been confirmed by outside testers.

Key facts

• Kimi says K3.1 Pro supports context of up to 1 million tokens.
• A faster version, K3.1 Pro Highspeed, runs two to three times as fast, according to Kimi.
• Both models are live today in the Kimi apps and through its API, the company says.
• Kimi says the full weights will be released by Nov. 4.
• All benchmark figures come from Kimi and have not been independently verified.

Background

K3.1 Pro is the first model in Kimi's K3.1 series. The company says it is built on Kimi K3 with additional post-training that received much more compute during post-training, the stage where a model is refined after its initial training.

Kimi's developer documentation lists the older K3 as an open-source, 3T-class model containing 2.8 trillion parameters. Kimi has not given an exact parameter count for K3.1 Pro. It describes the new model only as "3T-class."

The company has also been clearing out older models. The page now points users of those models to kimi-k3.1-pro.

What Kimi says the model can do

Kimi's model list describes kimi-k3.1-pro as having built-in visual understanding along with the 1 million-token window. The company also says the model can work with game engines including Unity, Unreal Engine and Godot. Kimi adds that a feature called 3D Make is now live on kimi.ai and in the Kimi app.

Kimi published several benchmark results. All of them are the company's own figures:

• Terminal-Bench 4.0: Kimi reports K3.1 Pro scored 65.7%, above GPT-6 Astra at 58.2% and just below Claude Opus 5.5 at 66.4%. Outside testers have not confirmed this result.
• Kimi Code Bench 3.0 Long-horizon: On this internal test, Kimi says the model reached 61.2%, up from 55.95% for K3.
• BenchCAD with tools: Kimi reports 92.7% for K3.1 Pro, compared with 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra.
• Hallu Bench Hard: Kimi says its rate of false claims on this internal stress test dropped from 23.4% with K3 to 8.3%.

Safety changes for the open release

Kimi says it took steps to limit how the public version could be misused for hacking. According to the company, it stripped cybersecurity-related material out of the training data for the open-weight release.

Kimi says it measured the effect on its own exploit test, called xdev-bench-hard. The final model solved 81.0% of tasks at pass@3, Kimi reports, compared with 90.5% for an earlier checkpoint built without those safeguards. Pass@3 means the model gets three tries at each task.

What it means for you

For regular users, K3.1 Pro is already available in Kimi's consumer products. Kimi says both new models can be used on Kimi.ai, Kimi Work and Kimi Code.

For developers, the bigger change may be the retirement of older models. Kimi's documentation also says the company plans to phase out every K2 model by late December 2026.

Researchers and companies that want to run the model on their own hardware will need to wait for the weights. The license terms for those weights have not been announced.

Practical steps

• Try it in the apps: Kimi says K3.1 Pro is live on Kimi.ai, Kimi Work and Kimi Code.
• Use the API: Developers can call the models as kimi-k3.1-pro and kimi-k3.1-pro-highspeed, according to Kimi's model list.
• Plan a move off K2: Kimi says it will end K2 support in late December 2026, so apps that rely on them will need to switch.
• Watch for the weights: Kimi says the full weights will arrive by Nov. 4.

What happens next

The next milestone is the open-weight release. Kimi says it will publish the full weights by Nov. 4. Once the weights are public, outside researchers will be able to test the model themselves. So far, none of the allowed news outlets have covered it, and the benchmark claims remain Kimi's alone.

FAQ

Can I download Kimi K3.1 Pro now? No. Kimi says the full weights will be released by Nov. 4.

How much text can K3.1 Pro handle at once? Kimi says it supports a context window of 1 million tokens.

What is K3.1 Pro Highspeed? It is a faster version of the model. Kimi says it runs two to three times as fast as K3.1 Pro.

Are the benchmark results verified? No. All the scores come from Kimi, and outside testers have not confirmed them.

What happens to the K2 models? Kimi says new users lose access to K2 models on Oct. The company says the full K2 lineup is due to be retired in late December.

Can K3.1 Pro work with game engines? Kimi says the model supports Unity, Unreal Engine and Godot. The company also says its 3D Make tool is available on kimi.ai and in the Kimi app.

Did Kimi change the model for safety? Kimi says it took cyber-related data out of training for the open release, which lowered the model's score on its own exploit test.

What license will the weights use? Kimi has not announced the license terms.