Chinese AI startup Moonshot AI is set to release a large language model that matches the performance of U.S. frontier models.

The Financial Times (FT) reported on July 16, citing two sources familiar with the matter, that Moonshot AI will release China’s largest AI model, Kimi K3, within days.

Kimi K3 is expected to have 2 to 3 trillion parameters, it said. The number of parameters refers to the size of a model’s neural network, and higher numbers tend to indicate stronger performance.

Anthropic has not disclosed parameter counts, but industry experts estimate Claude Opus 4.8 to be about 1.5 to 2 trillion parameters.

Sources said K3 is expected to surpass Opus 4.8 on major benchmarks and be available for free download as an open-weight model. They said it is expected to fall short of Anthropic Fable.

The K3 launch is drawing attention because it challenges the industry belief that Chinese AI models lag U.S. performance by 8 to 12 months. The fact that anyone can download it for free is a major challenge to Anthropic and OpenAI, which stick to expensive closed models, Reuters reported.

The latest U.S. models still lead Chinese models in complex tasks, but more U.S. investors and executives see the gap narrowing. Marc Andreessen (마크 앤드리슨), co-founder of Andreessen Horowitz, last month called Z.ai’s GLM-5.2 "the first Chinese model that is comparable to, or often surpasses, open models from major U.S. labs."