Moonshot AI releases Kimi K3, a 2.8T-parameter MoE model

Moonshot AI has published Kimi K3, the latest entry in its Kimi family, releasing the weights openly on Hugging Face. The model is a mixture-of-experts (MoE) system with roughly 2.8 trillion total parameters, positioning it among the largest openly distributed models to date.

K3 is described as multimodal, handling both text and vision inputs, with an emphasis on reasoning. Moonshot points to coding and agentic workloads as areas of particular strength — the kinds of tasks that increasingly define how frontier-scale models are judged, from writing and debugging software to orchestrating multi-step tool use.

Why it matters

Openly released weights at this scale remain rare. Most models in the multi-trillion-parameter range stay behind APIs, so a downloadable MoE of this size gives researchers and companies a chance to study and self-host capabilities that are usually gated.

• Architecture: mixture-of-experts, ~2.8T total parameters
• Modalities: text and vision, with a reasoning focus
• Focus areas: coding and agentic performance
• Access: weights available on Hugging Face under a custom license

A few practical details, including context length and the exact license terms, aren't spelled out in the release record, so teams evaluating K3 for production should check the model card directly. As with earlier Kimi releases, the real test will be independent benchmarking and how the model holds up on long-horizon agentic tasks outside curated demos.