Kimi K3.1 Pro

What it is

Kimi K3.1 Pro is a large language model released by Moonshot AI in October 2026. Moonshot presents it as an upgrade aimed at financial research, 3D modeling and coding ^[1][2]. It is the first model in the K3.1 series, and its weights are published openly under a custom license, the Kimi K3.1 Pro License ^[3].

How it was built

Kimi K3.1 Pro was post-trained from Kimi K3 with large-scale joint reinforcement learning ^[3]. Post-training is the stage that follows the bulk of a model's training on text: the model is pushed toward particular behaviours, such as following instructions, calling tools or working through a problem step by step. Reinforcement learning is one way of doing that — the model produces attempts, the attempts are scored, and the model is adjusted so that higher-scoring attempts become more likely. "Joint" here refers to training several kinds of task together rather than one at a time.

Scale and architecture

The model keeps the scale of Kimi K3. It has 2.8 trillion total parameters, of which 104 billion are active at a time, spread across 896 experts, with 16 experts selected and 2 shared for each token ^[3]. That arrangement is a mixture-of-experts design: instead of running every parameter for every token, a routing step picks a small subset of specialised sub-networks, so the total parameter count can be very large while the computation per token stays closer to that of a much smaller model.

It also keeps a 1-million-token context window and a MoonViT-V2 vision encoder, which is what lets the model take images as well as text ^[3].

The main architectural change is in attention, the part of a transformer that lets each token look at other tokens. Kimi K3 used full-attention Gated MLA layers; K3.1 Pro replaces them with Gated DeepSeek Sparse Attention (Gated DSA) layers, giving a stack of 69 KDA layers plus 24 Gated DSA layers ^[3]. Sparse attention means a token attends to a selected subset of the other tokens rather than to all of them, which keeps the cost of long inputs from growing as quickly as it does under full attention — a relevant property for a model built around a very long context window.

Reported results

The published scores are Moonshot's own. On BrowseComp the model scores 92.1, on HLE-Full with tools 70.0, on GPQA-Diamond 94.2, on OSWorld V2.1 79.6 and on DeepSWE v1.1 75.5 ^[3]. Moonshot also reported Terminal-Bench 4.0 at 65.7, SWE Marathon v1.1 at 51.7, and MMMU-Pro at 90.0 with Python ^[3]. On GDPval-AA v2.1 the reported figure is an Elo of 1846 ^[3]. The model card gives the context window as 1,048,576 tokens ^[3].

How to use it

Kimi K3.1 Pro is available through the Kimi API under the name kimi-k3.1-pro, with endpoints compatible with the OpenAI and Anthropic formats, so existing client code can point at it with relatively little change ^[3]. It also exposes a reasoning effort setting with three levels — low, high and max — which trades speed against how much internal computation the model spends before answering ^[3]. The model card directs users to platform.kimi.ai to select kimi-k3.1-pro ^[3].

Where it sits

K3.1 Pro is an incremental step rather than a new generation: the same parameter count, the same context length and the same vision encoder as K3, with retraining and a different attention stack on top ^[3]. Its release as open weights under a custom license makes the model available for download as well as through the API, and its stated targets — financial research, 3D modeling and coding — place it among general-purpose models positioned for long, tool-heavy professional work rather than chat alone ^[1][2][3].