Moonshot AI has published the full weights of Kimi K3 on Hugging Face, meeting the July 27 deadline the company set at launch. The 1.56-terabyte release makes the largest open-weight model in history freely downloadable, accompanied by a technical report describing how the 2.8-trillion-parameter system was trained. The mixture-of-experts model supports a 1-million-token context window and native visual understanding, and posted near-frontier results against Claude Fable 5 and GPT-5.6 Sol in company-run benchmarks at launch.
The license differs from what Moonshot indicated at launch. The company had pointed to a modified MIT license; the weights shipped under a custom Kimi K3 License, MIT-derived and permissive for most uses, but requiring a separate agreement with Moonshot for model-as-a-service businesses whose aggregate revenue from the model exceeds $20 million over any 12-month period. The release topped Hugging Face’s trending chart with more than 4,000 likes in its first half hour, reported as the fastest release growth in the platform’s history, and mirrors and quantized variants began appearing within hours. Hosted access spread as quickly: within a day OpenRouter listed K3 from seven providers, most of them matching Moonshot’s own pricing of $3 per million input tokens and $15 per million output tokens.
The release arrived while Washington was still debating restrictions. In the eleven days between K3’s launch and the weights publication, the White House accused Moonshot of building the model by distilling Anthropic’s Fable and the administration weighed restrictions on Chinese models, with analysts noting that downloadable weights cannot be recalled or geofenced once published. Any US measures adopted now would apply to a model that has already been downloaded and mirrored worldwide.
The technical report documents training methods that frontier labs typically treat as proprietary, including the quantization-aware training that allows the model to ship at four-bit precision.
Running the model requires substantial hardware. Even at four-bit precision the weights occupy 1.56 terabytes, against roughly 5.6 terabytes at 16-bit, and loading the model takes around eighteen 80-gigabyte accelerators, or eight of the newest 192-gigabyte cards. That is the same constraint that forced Moonshot to pause new consumer subscriptions when demand outran its own capacity. Organizations that download the weights take on that hardware cost themselves, in most cases by renting the compute from a cloud provider.
Beijing is separately weighing export restrictions of its own, including limits on foreign downloads of Chinese model weights. K3’s weights were published before any such rule took effect.
Sources: Hugging Face, Simon Willison, OfficeChai
–
By the Control Plane Editorial Team