Alibaba has introduced Qwen3, a new family of AI reasoning models that the company claims can compete with and even surpass top models from Google and OpenAI. The Qwen3 models range from 0.6 billion to 235 billion parameters, with most being available for download under an open license on platforms like Hugging Face and GitHub. These models are designed as hybrid systems capable of both quick responses and complex reasoning, which allows them to fact-check themselves effectively.
Qwen3 incorporates a mixture of experts (MoE) architecture for improved computational efficiency, supporting 119 languages and trained on a dataset of nearly 36 trillion tokens. While Qwen3 models do not consistently outperform the latest models from competitors, they show strong performance in various benchmarks, such as coding and math tasks. The flagship model, Qwen-3-235B-A22B, has demonstrated superior results in specific programming contests and math evaluations, although it is not yet publicly available.
Alibaba emphasizes the tool-calling capabilities and instruction-following features of Qwen3, which are further enhanced by its integration of thinking and non-thinking modes. As competition intensifies, models like Qwen3 exert pressure on American AI labs to innovate, while U.S. policies restrict Chinese access to necessary training chips. Overall, Qwen3 represents a significant advancement in AI capabilities from Alibaba.
