Kimi K3.1 Pro brings a 3T-class open-weight model to users

Moonshot AI has released Kimi K3.1 Pro, the first model in its new K3.1 series. In its technical announcement, the company describes it as a post-trained version of Kimi K3 with 2.8 trillion total parameters and 104 billion active parameters in a mixture-of-experts design. It has a context window of up to one million tokens and supports multimodal input. Moonshot also launched Kimi K3.1 Pro Highspeed, which it says produces responses about two to three times faster while keeping the same model quality.

The model is available immediately through Kimi.com, Kimi Work, Kimi Code and the Kimi API. Moonshot says the complete model weights will be released by 4 November 2026. That makes the current launch a hosted release first, while the promised weights could later give developers more control over deployment, evaluation and adaptation. A model of this size will still require substantial infrastructure, even though the mixture-of-experts design limits the parameters used for each token.

Moonshot presents K3.1 Pro as a long-horizon work model. Its examples cover full-project coding with Unity, Unreal Engine and Godot, as well as parametric modelling in Blender and CAD tools. The company reports 65.7 percent on Terminal-Bench 4.0, compared with 58.2 percent for GPT-6 Astra and 66.4 percent for Claude Opus 5.5. On BenchCAD with tools, it reports 92.7 percent. These are vendor-reported results, so they are useful evidence about the company’s targets but not an independent ranking.

The announcement also includes a safety choice for the open-weight release. Moonshot says it removed cyber-related data from mid-training and reinforcement-learning stages, while retaining general coding data. On its internal xdev-bench-hard test, the resulting model scored 81.0 percent pass@3, below the 90.5 percent score of a comparison checkpoint without those measures. That does not eliminate misuse risk, but it shows the lab is treating the open release as a separate safety decision.

For users and makers, the practical news is the combination of very long context, multimodal work and a promised open release at a scale rarely available outside a few labs. The important follow-up will be whether the weights arrive on time, under what licence, and how independent tests measure quality, cost and safety. Until then, K3.1 Pro is a significant new hosted option, not yet a locally deployable model.