China's Open Source AI Models Ranked World's Best

Chinese open-source AI models have been recognized as the best in the world, ahead of developments by giants such as Google and Meta*. According to the rating of the LMArena platform, created by researchers at the University of California, Berkeley, the first places were taken by China's Kimi K2, MiniMax M1, Qwen 3 and DeepSeek R1. These models were rated as superior in response quality and logic to such well-known Western counterparts as Google's Gemma 3-72B and Meta's Llama 4-Maverick*.

The leader of the rating is the Kimi K2 model, developed by Moonshot AI. It was presented on July 11, and is already considered one of the most impressive open large language models. Users note that its answers are distinguished by humor and naturalness, which makes it especially popular. The second place was taken by DeepSeek R1-0528 from the Hangzhou startup DeepSeek, which copes well with multi-stage dialogues and complex logical tasks.

The success of these models confirms that Chinese companies are actively closing the gap with American leaders in the field of AI, relying on open source code. For example, DeepSeek released a powerful language model V3 in December 2023, and in January it presented R1, which is comparable in capabilities to OpenAI developments.

Another key player was Alibaba Group Holding with the Qwen 3, which took third place in LMArena's ranking for "pure reasoning power." According to Hugging Face, three of the four Chinese models in the top 10 are from this series.

China’s progress in open-source AI has even been praised by Nvidia CEO Jensen Huang. During a recent visit to China, he called DeepSeek, Qwen, and Kimi the world’s most advanced models, highlighting their leadership in inference. With U.S.-China trade relations improving, Nvidia is preparing to resume shipments of its cutting-edge H20 chips to China, which could give a new boost to local AI development.