Moonshot AI says Kimi K3 can challenge OpenAI and Anthropic

Moonshot AI has unveiled Kimi K3, a new model the Chinese start-up says can compete with top systems from OpenAI and Anthropic. The model contains 2.8 trillion parameters, a measure of scale and processing power, and is expected to reveal its full capabilities in coding, knowledge work and reasoning when it is released as an open-source model on 27 July.

The company said Kimi K3 is its “most capable flagship model to date” and is designed to operate with “minimal human supervision” on tasks such as engineering and coding. Unlike closed US systems, its open-source design would allow outside developers to download, run, customize and adapt it for advanced reasoning and complex software work.

Independent evaluations from Artificial Analysis and Arena.ai showed Kimi K3 performing on a par with leading US models, including OpenAI’s GPT and Anthropic’s Claude. It ranked first in web interface engineering and beat Anthropic’s Fable system in blind human-preference tests.

The launch comes as the US treats advanced AI software as critical national infrastructure and frontier models as national security assets. Shares in Chinese rivals Zhipu and MiniMax fell sharply in Hong Kong after the announcement, by about 27% and 16% respectively.