Moonshot AI released Kimi K3, a 2.8 trillion parameter mixture-of-experts model that activates 104 billion parameters per token, according to the model's GitHub repository.
The model has 93 layers split between 69 KDA attention layers and 24 gated MLA attention layers, routes among 896 experts with 16 active per token, and supports a 1 million token context window with native text, image and video input, the repository says.
Moonshot reported benchmark scores including 93.5% on GPQA Diamond, 67.5% on DeepSWE, 88.3% on Terminal-Bench and 91.2% on BrowseComp, and said the model performs competitively with Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5 on several coding and reasoning tests.
The weights are released under a custom Kimi K3 License covering research, deployment and further development, rather than a standard permissive open source license, the repository says.
Within days of the release, an independent developer published a C inference engine that runs the full model on a single CPU using about 8.2 gigabytes of peak memory, down from the shipped checkpoint's 1.56 terabytes, by loading only the small number of active experts each token needs and streaming dense layers from disk, according to that project's own repository.
For builders, the model's scale needs serious infrastructure to run as shipped, but that inference engine shows how fast independent developers can shrink that requirement: what needed a server rack on release day needed one CPU and a spare hard drive within about a week. The license terms, not the hardware, look like the bigger constraint on what anyone can build with it.