Moonshot AI releases full model weights of its Kimi K3, world’s largest open-weight AI model
China’s Moonshot AI has officially released the full model weights of its Kimi K3, a 2.8-trillion-parameter system, on Hugging Face, which is currently the world’s largest open-weight AI model, available for free download, fine-tuning, and self-deployment.
The model features a 1-million-token context window and adopts a Mixture-of-Experts architecture, activating approximately 50 billion parameters per token during computation.
Unlike closed models that are only accessible via paid APIs (Application Programming Interface), an open-weight model allows anyone to download and run the trained AI system on their own infrastructure.
Demand surged within 48 hours, putting pressure on Moonshot’s computing capacity, prompting the company to suspend new user subscriptions and prioritize service quality for existing paying users. The company offers K3’s API at US$3 per million input tokens and US$15 per million output tokens, roughly half that of some U.S. competitors.
Moonshot said K3 is the first open-weight model to approach the 3-trillion-parameter threshold, with competitive performance across multiple benchmarks. The release puts pressure on closed-source players like OpenAI and Anthropic, as developers can now download, fine-tune, and deploy the model without being locked into proprietary APIs.
However, open weights do not mean low barriers to self-hosting. Reuters previously reported that most users are unlikely to deploy such a large model on their own due to extremely high hardware costs. In other words, while K3 is more open in its business model, the organizations truly capable of fully leveraging its capabilities remain those with large-scale computing resources.
Moonshot is also reportedly in talks with Goldman Sachs regarding a Hong Kong initial public offering, as Chinese AI labs continue to challenge U.S. dominance through open-weight strategies.
