Moonshot AI open-sources the trillion-parameter Kimi K2

Background

Moonshot AI is a Chinese AI company, and Kimi K2 was one of its large language models. It was released as open weights in July 2025, meaning the trained weights were published so that others could download, run and study the model ^[1][2].

Parameters are the numerical values a model learns during training, and the total count is a rough measure of how much a model can store.

What happened

In July 2025 Moonshot AI released Kimi K2 as an open-weight language model ^[1]. It is a mixture-of-experts model with 1 trillion total parameters, of which 32 billion are active for each token ^[1][3]. The base model was pre-trained on 15.5 trillion tokens. Moonshot described the training as having zero training instability, and the pre-training used its MuonClip optimizer, which applies the Muon optimizer at unprecedented scale ^[3][1].

Key details

The model has 61 layers, one of them dense, and 384 experts. For each token, 8 experts are selected by a router, and 1 shared expert is always used. Attention is implemented as Multi-head Latent Attention, with 64 heads ^[3].

The weights are released under a modified MIT license ^[3].

Moonshot's reported results for the instruct version, Kimi-K2-Instruct, include 65.8 on SWE-bench Verified in a single attempt and 71.6 when multiple attempts are allowed. SWE-bench Verified tests whether a model can fix real issues in open-source software repositories. The model also scored 53.7 on LiveCodeBench v6, which tests programming problems, and 49.5 on AIME 2025, a set of competition mathematics problems ^[3].

Impact

Kimi K2 showed that a Chinese lab could publish a model of trillion-parameter scale with open weights. For developers and researchers the release offered a very large model that could be studied, tested and built on, rather than only accessed through an interface.