Moonshot AI Releases Kimi K3, Its Largest Open-Weight Model

Chinese AI lab Moonshot AI has published the complete open weights for Kimi K3, a 2.8-trillion-parameter model the company calls the first open model in the 3-trillion-parameter class. The weights went live on July 26, a day ahead of Moonshot’s own schedule, alongside a technical report and inference code.

What’s new in K3

Kimi K3 uses a sparse mixture-of-experts design that activates only 16 of its 896 experts for any given token, which keeps inference costs down despite the model’s overall size. Moonshot combined that with two attention components it built in-house, Kimi Delta Attention and Attention Residuals, spread across 93 layers. The company says the architecture delivers roughly 2.5 times the scaling efficiency of its predecessor, Kimi K2.

The model handles a 1-million-token context window and processes text, images and video natively. To make a model this size practical to run, Moonshot trained it with quantization built into the process, shipping weights in a compact MXFP4 format instead of compressing a full-precision model after the fact.

The release includes the weights, the technical report and support for inference engines such as vLLM and SGLang, all under Moonshot’s own Kimi K3 license, which allows research and commercial use. The company also open-sourced some of the infrastructure behind K3, including its attention kernels and mixture-of-experts communication library.

Where it lands competitively

Moonshot said in its announcement that K3’s overall performance “still trails” Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol, but that it outperforms other models it tested on the same evaluation suite, particularly on coding and agentic tasks. The model is also available as a hosted service through Moonshot’s Kimi app and API, priced at $0.30 per million input tokens on a cache hit, $3 per million on a cache miss, and $15 per million output tokens.

The release adds to a broader push toward large open-weight models out of Chinese labs, including Nvidia chief Jensen Huang’s public call for the industry to back open models. It also comes months after the White House accused Moonshot of training K2 on outputs distilled from Anthropic’s Claude models, a claim Moonshot denied. Publishing K3’s full weights gives outside researchers and developers a way to inspect and run the model directly, rather than relying solely on API access.