Alibaba introduces new Qwen3 AI models with "reasoning"

Chinese tech giant Alibaba has unveiled new Qwen3 AI models. The company claims that these models can match or even surpass the latest developments from Google and OpenAI. The models have between 0.6 and 235 billion parameters and are available for download on Hugging Face and GitHub.

Qwen3 are "hybrid" models that combine speed for simple tasks with the ability for complex "thinking" for fact-checking. Some models use a Mixture of Experts architecture to improve efficiency.

The models support 119 languages and were trained on textbook data, code, and content generated by previous models. The new generation of Qwen shows significant improvements in tests. While none of the Qwen3 models outperform OpenAI’s latest developments, such as o3 and o4-mini, they still show strong results, notably outperforming OpenAI o1 in some tests.

According to Alibaba, Qwen3 stands out for its ability to accurately follow instructions. By the way, experts note that Qwen3, like other open models, could become an important tool for China against the background of restrictions on the import of chips for AI training.