Kimi K3 puts China back in the AI race

Moonshot AI’s Kimi K3 has put China back at the center of competition in generative AI, with public benchmark results placing it ahead of Anthropic’s Claude Fable 5 and OpenAI’s GPT5.6 Sol. Scheduled for release on July 27, the open-source model will allow researchers, institutions and companies to inspect, modify and adapt its code.

Kimi K3 was developed with 2.8 trillion parameters, compared with 1.6 trillion for the largest DeepSeek model. It processes one million tokens in a single task and has shown particular strength in programming, ranking first in the LLM Arena leaderboard for creating online interfaces and generating a playable 3D environment from a single photograph.

Moonshot AI says the model activates only the parameter subsets needed for each query, an efficiency approach intended to reduce unnecessary computation and energy use. Its debut follows DeepSeek’s January 2025 breakthrough and comes despite U.S. restrictions on selling Nvidia chips to China.

The launch also coincided with China’s World Artificial Intelligence Cooperation Organization, an AI governance initiative that includes about 30 countries, including Russia, Pakistan, Indonesia, Brazil, Cuba and Venezuela. Chinese President Xi Jinping described it as a response to calls from the Global South to promote AI development and governance.