China releases the world's largest AI model – free to download
Moonshot AI's Kimi K3 packs 2.8 trillion parameters and lands just behind the top models from Anthropic and OpenAI. The White House accuses the company of chip smuggling and model copying.
In the early hours of Monday CET – a full day ahead of the company’s own announced date – Chinese AI lab Moonshot AI published the model weights of Kimi K3 for free download. At 2.8 trillion parameters it is by far the largest open-weight AI model ever released, and the first to sit directly behind the closed frontier models in performance.
On the independent Artificial Analysis Intelligence Index, K3 debuts in third place, behind only Anthropic’s Claude Fable and OpenAI’s GPT-5.6 Sol Max. On Frontend Code Arena, a benchmark testing web development, it tops the list outright. No open model has ever been this close to the frontier.
1.4 terabytes
Free does not mean easy to use. The model files weigh in at 1.4 terabytes in compressed four-bit form, and simply loading the model requires roughly 18 advanced accelerators with 80 gigabytes of memory each. In practice, only data centres and large cloud providers can run it – hosting providers Together AI and Modal offered K3 from day one.
The model uses a mixture-of-experts architecture: of 896 expert modules, only 16 activate for each generated token, about 50 billion parameters at a time. It also handles a context window of one million tokens – the equivalent of several thousand pages of text in a single pass.
Accusations from the White House
The launch lands in an increasingly sharp AI rivalry between the US and China. Michael Kratsios, director of the White House Office of Science and Technology Policy, last week accused Moonshot of training K3 on export-restricted Nvidia chips and of large-scale "distillation" – using American models’ outputs to train its own – including against Anthropic’s Fable model. Moonshot has denied the accusations, which have not been publicly documented.
Meanwhile, China’s president Xi Jinping has publicly committed the country to open-source AI at the World AI Conference, and Alibaba has announced that its next large model, Qwen 3.8 at 2.4 trillion parameters, will also get open weights. The gap between open and closed models has shrunk from six–nine months to three–five months in a year, according to analysts.
Cheaper AI for everyone
For ordinary users and businesses, the main consequence is competition: when a near-frontier model can be hosted by anyone with the hardware, prices for AI services fall across the market. Several European hosting providers are expected to offer K3 from their own data centres, which could also make advanced AI available without data leaving Europe.
