Moonshot Releases Kimi K3’s 2.8T Open Weights
Moonshot AI has released the full weights of Kimi K3, giving developers access to a 2.8-trillion-parameter multimodal model they can modify and self-host.
One of the largest publicly downloadable AI models is now open for developers to inspect, modify and run on their own infrastructure.
What happened
Moonshot AI released the full weights of Kimi K3. The mixture-of-experts model has 2.8 trillion total parameters, with 104 billion active during inference, alongside native multimodal capabilities and a one-million-token context window.
Unlike a model available only through an API, an open-weight release can be downloaded, adapted and self-hosted.
Why it matters
Open weights give companies more control over privacy, deployment and customisation. Developers can study how the model behaves, fine-tune it for specialised tasks and avoid depending entirely on one hosted provider.
The trade-off is infrastructure. A model of this size is expensive and technically demanding to run, so practical access may still be concentrated among well-funded teams and cloud providers.
The bigger picture
Kimi K3 adds pressure to closed-model companies by narrowing the gap between downloadable systems and proprietary frontier models. It also strengthens China’s role in the open-weight AI ecosystem.
Moonshot’s benchmark and efficiency claims remain company-reported, so the release should be judged through independent testing rather than headline parameter counts alone. The important signal is that increasingly capable models are becoming available outside closed APIs.