The Chinese technology company Alibaba on Monday unveiled the family of artificial intelligence models Qwen3. The models are available for download under an open license on the Hugging Face and GitHub platforms, their size ranges from 0.6 billion to 235 billion parameters. "We seamlessly integrated thinking and non-thinking modes, offering users flexibility in managing the thinking budget," wrote the Qwen team in a blog post. The models support 119 languages and were trained on a dataset of nearly 36 trillion tokens.
Alibaba claims significant improvements for the Qwen3 model compared to its previous version, Qwen2. The largest model, Qwen-3-235B-A22B, showed the best results on the Codeforces platform, surpassing OpenAI's o3-mini and Google's Gemini 2.5 Pro. The model demonstrates higher performance in AIME and BFCL tests compared to o3-mini. The public model Qwen3-32B is competitive with many proprietary and open AI models, including DeepSeek's R1. Alibaba claims that Qwen3 excels in tool invocation, instruction following, and data format copying. The model is available through cloud providers such as Fireworks AI and Hyperbolic.