Chinese artificial intelligence startup Moonshot AI has introduced a new large-scale AI model, Kimi K3, claiming it can compete with leading US models developed by OpenAI and Anthropic.
The company said Kimi K3 has 2.8 trillion parameters, a key measure of an AI model's size and computing capability.
Its full features, including coding, reasoning and knowledge-based tasks, will become available when the model is released as open source on July 27.
If launched as planned, Kimi K3 will become the world's first freely downloadable and customizable open-source AI model in the three-trillion-parameter class, allowing developers worldwide to use and modify it.
The launch signals China's rapid progress in artificial intelligence, narrowing the technology gap with US companies and challenging the long-held belief that Chinese AI developers lag behind their American counterparts.
The announcement comes weeks after the US government temporarily ordered American AI company Anthropic to withdraw its flagship Fable and Mythos models over cybersecurity concerns. Although those restrictions have since been lifted, Washington has increasingly treated advanced AI systems as critical national security assets and imposed tighter export controls on related technologies.
Despite US restrictions on advanced chip exports, Moonshot's latest model suggests Chinese AI companies continue to make significant progress using domestic resources.
Supported by Chinese technology giants Alibaba and Tencent, Moonshot AI has quickly emerged as one of the country's leading developers of generative AI.
In a statement, the company described Kimi K3 as its "most capable flagship model" so far.
Unlike the closed AI systems offered by OpenAI and Anthropic, Kimi K3 will be open source, enabling developers to customize it for complex software development, reasoning and other advanced applications.
Moonshot said the model is designed to perform engineering, coding and other complicated tasks with minimal human supervision.
Independent evaluations by Artificial Analysis and Arena.ai found that Kimi K3 performs at a level similar to leading US AI models, including OpenAI's GPT and Anthropic's Claude.
According to the benchmark results, Kimi K3 ranked first in web interface engineering and outperformed Anthropic's Fable model in blind human preference tests.
Although the model requires powerful computing hardware to run locally because of its massive size, analysts say its open-source release could reshape the AI market by giving developers free access to advanced technology.
Following the announcement, shares of Moonshot's Chinese rivals Zhipu and MiniMax fell sharply in Hong Kong, dropping about 27% and 16%, respectively.
Source: BBC