Chinese Open-source AI Models Occupy Top Spots Among Global Developers: Ranking

A new wave of Chinese open-source AI models is reshaping the global leaderboard, challenging the dominance of American tech giants like Meta and Google. According to LMArena, a widely used benchmarking platform developed by researchers at UC Berkeley, China now leads the pack in open-sourced large language models (LLMs), with several homegrown models outperforming major Western contenders in direct, response-based comparisons.

In its most recent report, LMArena named Kimi K2, MiniMax M1, Qwen 3, and a version of DeepSeek R1 as the top open-source LLMs globally outshining high-profile entries such as Meta’s Llama 4-Maverick and Google’s Gemma 3-72B. LMArena, formerly known as Chatbot Arena, determines its rankings through a crowdsourced evaluation method that pits AI model responses against each other, with users voting on the more effective or natural-sounding output.

The top spot went to Kimi K2, launched by Shanghai-based startup Moonshot AI on July 11. LMArena praised the model for being “one of the most impressive” open-sourced LLMs released to date, specifically highlighting its “humorous without sounding too robotic” conversational tone. Moonshot has open-sourced two versions of Kimi K2, increasing its accessibility and appeal in the AI community.

Coming in second was DeepSeek R1-0528, a refined iteration of the base R1 model released in January by DeepSeek, another Chinese startup based in Hangzhou. This version excelled in multi-turn conversations and reasoning tasks, areas that are often seen as key benchmarks for advanced AI performance.

Alibaba’s Qwen 3 series also made a powerful showing, with its 235-billion-parameter version taking third place. LMArena cited its “raw reasoning power” as a distinguishing factor. The Qwen series has also been dominant on Hugging Face’s open LLM leaderboard, with three Qwen models ranked in the global top 10 as of last month.

This rise of Chinese open-source AI is significant not just in terms of performance, but in its philosophy and accessibility. Unlike closed systems developed by some Western firms, these models are released with open code and model weights, making them free to use, adapt, and deploy. According to LMArena, this open-source momentum has “sent shock waves through Silicon Valley and Wall Street”, especially after the debut of DeepSeek’s free-to-use V3 and the launch of R1 in late 2024.

The success has also caught the attention of Nvidia CEO Jensen Huang, who during a visit to China last week, stated that Kimi, DeepSeek, and Qwen were currently the “best open reasoning models in the world today.” Huang emphasized that China’s models were now “very advanced” and comparable with, if not ahead of their Western counterparts.