Chinese AI models topped the list of open-source models in a ranking tool developed by UC Berkeley researchers, in a development that shows how the open-source movement is helping Chinese models gain on Western rivals.
A number of Chinese AI companies have begun open-sourcing AI models since January, when Hangzhou-based start-up DeepSeek gained worldwide recognition for its high-performance, low-cost approach.
According to new rankings from LMArena, a ranking platform developed by UC Berkeley researchers, Moonshot AI’s Kimi K2, DeepSeek’s R1, Alibaba’s Qwen 3 and MiniMax’s M1 ranked at the top of the open-source AI listings.
Open-source strategy
Alibaba is one of China’s biggest tech companies, while Moonshot, DeepSeek and MiniMax are start-ups.
The models were more popular than open-source offerings from the likes of Google and Meta, although proprietary tools from Google, OpenAI and xAI ranked ahead of the open-source models.
The models were scored by allowing users to view results from multiple models side-by-side and then casting votes on the results, LMArena said in a report.
LMArena highlighted how Moonshot’s Kimi K2 had risen to the top of the open-source rankings after being released on 11 July.
LMArena said Kimi K2 was one of the most impressive open-source large-language models (LLMs) to date, and was gaining popularity because its chatbot responses were “humorous without sounding too robotic”.
DeepSeek R1-0528, came in second and was “strong in multi-turn dialogue and reasoning tasks”, the platform’s researchers said.
Alibaba’s Qwen 3-235b-a22b-no-thinking, launched in April with some 235 billion total parameters, ranked third for its “raw reasoning power”, LMArena said.
Start-ups
The Qwen series has also proven popular in separate rankings by the Hugging Face AI platform in June, where three of the four top 10 ranked Chinese LLMs were Qwen models.
The M1 model from MiniMax ranked fourth in the open-source listings.
MiniMax filed confidentially for a Hong Kong initial public offering, targeting a value of $4 billion (£2.9bn) for the flotation, which could happen before the end of this year, Reuters reported last week.
OpenAI has spoken of releasing an open-source model this year, but last week indefinitely delayed the project for further testing.
