Kimi K2

What it is

Kimi K2 is an open-weight language model released by Moonshot AI in July 2025 ^[1]. It is a mixture-of-experts model with 1 trillion total parameters and 32 billion active for each token ^[1][3]. In practice this means the model stores a very large body of knowledge across its parameters, while each token passes through only a fraction of them. Its reported scores cover coding, math and tool-use tasks, which are the kinds of work where a model calls tools and carries out multi-step operations.

Origins

Kimi K2 is a large open-weight language model, released under a modified MIT license ^[3]. Its base model was pre-trained on 15.5 trillion tokens, and Moonshot described the training run as having zero training instability. Pre-training used MuonClip, an optimizer that applies the Muon optimizer at unprecedented scale ^[3][1].

Moonshot also wrote a scaling paper with UCLA on the Muon optimizer. According to Moonshot, that paper trained a 16-billion-parameter mixture-of-experts model with 3 billion active parameters at about twice the computational efficiency of AdamW ^[1].

Main features

Architecturally, the model has 61 layers, one of them dense, and 384 experts. Eight experts are selected per token, plus one shared expert that every token uses. It uses Multi-head Latent Attention (MLA) with 64 heads. The context window is 128K tokens, and the vocabulary has 160K tokens ^[3].

The reported benchmark results for Kimi-K2-Instruct are 65.8 on SWE-bench Verified in a single attempt and 71.6 with multiple attempts, 53.7 on LiveCodeBench v6 and 49.5 on AIME 2025. Other reported scores are 75.1 on GPQA-Diamond, 89.5 on MMLU, 81.1 on MMLU-Pro, and 70.6, 56.5 and 65.8 on the Tau2 retail, airline and telecom tasks ^[3].

On 9 September 2025 the company released Kimi-K2-Instruct-0905, which doubled the context window from 128K to 256K tokens ^[1][2].

Kimi models were also used by other companies. Kimi models were the basis for, or were used in post-training of, Cursor's Composer 2 and Thinking Machines Lab's Inkling ^[2].

Its place in development

Kimi K2 is a trillion-parameter general model that others can download and use, which set it apart from the smaller open releases in the company's earlier line ^[3]. It was followed in November 2025 by the reasoning version, Kimi K2 Thinking ^[1].