Kimi K3.1 Pro: Moonshot's new open-weight AI model ranks 4th

China's Moonshot AI has released Kimi K3.1 Pro, and it is one of the strongest models you will soon be able to download and run yourself. Moonshot announced the model on 23 October as its most capable "3T-class" model, with a 1-million-token context window, plus a faster K3.1 Pro Highspeed variant that it says runs two to three times quicker.

Where it ranks

Independent benchmarker Artificial Analysis scored K3.1 Pro 55 on its Intelligence Index, putting it 4th overall, behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Meta's Muse Wonder Full (56). That is 11 points above Kimi K3. Once the weights are published, it says K3.1 Pro will be the highest-scoring open-weights model, and the gap between the best open and closed models on its index will shrink from 12 points to 3.

The biggest jumps are in agentic work, where a model runs tools and multi-step tasks on its own. Artificial Analysis measured 67% on Terminal-Bench 4.0, the top score of any model it has tested there, up from 13% for K3. Gains elsewhere, such as Humanity's Last Exam (47% to 52%), are smaller.

According to Artificial Analysis, the model has 2.8 trillion total parameters with 104 billion active, accepts text and images, and costs $2 per million input tokens and $10 per million output tokens through Moonshot's API, a third cheaper than K3.

Hacking skills deliberately trimmed

The more unusual part of the launch is about security. Moonshot says it found cyber-related material inside its general training data and RL tasks, and that leaving it in made the model's vulnerability-research skills climb steeply. For the open-weight release, it removed cyber-related data from both training stages.

On Moonshot's own exploit benchmark, a checkpoint without those cuts solved 90.5% of tasks, while the released model solves 81%. Moonshot says that keeps its offensive capability behind frontier proprietary models that have been public for at least three months. These are the company's own numbers, so outside evaluation will matter. The timing is notable after a week in which AI-assisted bug hunting surfaced the critical PHP2Shell flaws.

Can you run it at home?

Not easily. Even with only 104 billion parameters active per token, a 2.8-trillion-parameter model needs server-class memory. Most people will use it through the Kimi apps or API, while smaller distilled or quantised versions from the community are what usually reach desktops. If you're new to the idea, read our explainer on what open-weight AI models are, and our guide to choosing a laptop for local AI for what home hardware can realistically handle.