Alibaba releases a new generation of open-source AI reasoning models.
The Chinese tech company Alibaba has announced the Qwen3 series: a range of AI models that, according to the company, can compete with, and sometimes outperform, the best models from OpenAI and Google.
From 0.6 to 235 Billion Parameters
The models in the Qwen 3 family vary in size, from 0.6 to 235 billion parameters. The more parameters, the better the model performs. Most will soon be available as open models on Hugging Face and GitHub. The hybrid models combine fast response times with in-depth reasoning and provide users with a “thinking budget”. This allows them to more easily configure task-specific budgets. Some models use a “Mixture of Experts” architecture (MoE) for more efficient task distribution.
Extensively Trained
Alibaba says in the announcement that Qwen3 supports 119 languages and has been trained on 36 trillion tokens from textbooks, Q&As, code snippets, and synthetic data. Qwen3 excels in following instructions, invoking tools, and copying data formats. The largest model (Qwen-3-235B-A22B) narrowly beats competitors like OpenAI’s o3-mini and Google’s Gemini 2.5 Pro on mathematical benchmarks, for example, but is not yet publicly available.
The emergence of model series developed in China, such as Qwen, is putting pressure on America to develop better AI technology. According to experts, these models will be widely deployed, especially in China, despite American export restrictions on chips.
Qwen3 is already available through the Qwen Chat web interface or via the mobile app.