Moonshot AI has released the full weights for Kimi K3, giving developers the freedom to download, modify, fine-tune, and host the model themselves. The move comes shortly after K3’s debut, when Moonshot’s benchmarks placed it close to Claude Fable 5 and GPT-5.6 Sol, and even ahead in a handful of tests. K3 has 2.8 trillion parameters and a 1-million-token context window.
Downloading K3 is easy, but running it is not
Open weights make the trained parameters that shape a model’s behavior publicly available. Companies can adapt K3 using private data, build specialized tools around it, or offer hosted versions under Moonshot’s license without relying entirely on its API. K3 is still not fully open source, however, since Moonshot has not released its complete training data or every part of the training process.
The model is available through Hugging Face, but it is far too large for an ordinary PC. Its mixture-of-experts design activates 104 billion parameters at a time, while the MXFP4 weights alone are expected to occupy roughly 1.4TB before accounting for runtime overhead. Moonshot says K3’s new architecture delivers about 2.5 times more intelligence per unit of compute than Kimi K2. Even so, self-hosting the model will require a serious multi-GPU server setup.
K3 is now caught in a US policy fight
The release arrives in the middle of a growing US debate over Chinese AI and open-weight models. White House officials have accused Moonshot of distilling Anthropic’s Fable model and training K3 using restricted Nvidia hardware. Moonshot has not publicly responded to those claims.
Meanwhile, Nvidia CEO Jensen Huang is backing a growing push in the US to protect open-weight AI. The number of companies supporting a letter published on July 24 reportedly doubled from 25 to 50 within day, adding names such as OpenAI, Google, AMD, Cisco, Cloudflare, GitHub, Block, and Ollama. The group is urging Washington not to place sweeping restrictions on downloadable models, arguing that open weights support innovation, cybersecurity research, and the ability to run AI on private infrastructure.
K3’s release gives developers a capable new model to work with, but its size will limit who can actually run it. Cloud providers and larger companies are likely to be the first to offer it at scale. Amazon and Anthropic have not signed the letter, showing that major technology companies still disagree over how freely advanced AI models should be released.