Moonshot AI releases Kimi K3.1 Pro

Background

Moonshot AI develops the Kimi family of models. K3.1 Pro is an upgrade within the K3 line rather than a new model built from scratch.^[1][2][3]

What happened

In October 2026 Moonshot AI released Kimi K3.1 Pro. Moonshot described it as an upgrade for financial research, 3D modeling and coding.^[1][2] It was the first model in the K3.1 series, and its weights were published as open weights under a custom Kimi K3.1 Pro License.^[3]

The model was post-trained from Kimi K3 using large-scale joint reinforcement learning. Post-training refers to the training that follows the initial learning from large amounts of text. Reinforcement learning adjusts a model according to reward signals from the tasks it practices.^[3]

Key details

The release changed the attention layout. The full-attention Gated MLA layers in K3 were replaced with Gated DeepSeek Sparse Attention (Gated DSA) layers, giving 69 KDA layers plus 24 Gated DSA layers.^[3] Sparse attention lets each token attend to a selected set of earlier tokens rather than all of them, which lowers the cost of very long inputs.

Moonshot reported self-reported scores of 92.1 on BrowseComp, 70.0 on HLE-Full with tools, 94.2 on GPQA-Diamond, 79.6 on OSWorld V2.1 and 75.5 on DeepSWE v1.1.^[3] Self-reported figures come from the developer's own tests, not from a third-party evaluation.

Developers can call the model through the Kimi API as kimi-k3.1-pro. The API has OpenAI-compatible and Anthropic-compatible endpoints, and the reasoning effort can be set to low, high or max.^[3] A compatible endpoint accepts requests written for another provider's interface, so existing code can often be reused with a change of address and model name.

Significance

The release shows the K3 design being extended in two ways at once: an updated attention layout and added training for particular kinds of work. Its weights are open, under a custom license. Readers who want the full architecture numbers should consult the model entry for K3.1 Pro.