Chinese technology alibaba on Monday released Qwen 3, a family of AI models, the company affirms corresponds to the matches and, in some cases, surpasses the best models available from Google and Openai.
Most models are – or will soon be – available for download under an “open” license from the AI Dev platform Face And Github. They vary in size from 0.6 billion parameters to 235 billion parameters. The parameters roughly correspond to the problem solving skills of a model, and the models with more parameters generally work better than those with fewer parameters.
The rise in the series of models from China as Qwen has increased pressure on American laboratories such as OpenAi to provide more competent AI technologies. They also led decision -makers to implement restrictions to limit the capacity of AI Chinese companies to obtain the chips necessary to form models.
According to Alibaba, Qwen 3 models are “hybrid” models in the sense that they can take time and “reason” through complex problems or quickly respond to simpler requests. The reasoning allows models to be checked effectively, similar to models like O3 of Openai, but at the cost of a higher latency.
“We have modes of thinking and non-thought integrated in a transparent way, offering users the flexibility to control the budget of thought,” wrote the Qwen team in a blog. “This design allows users to configure specific budgets to tasks with greater ease.”
The Qwen 3 models take charge of 119 languages, known as Alibaba, and were trained on a set of data of almost 36 billions of tokens. The tokens are the raw data bits that a model processes; 1 million tokens is equivalent to around 750,000 words. Alibaba says that Qwen 3 was formed on a combination of manuals, “pairs of answers to questions”, code extracts, data generated by AI, and more.
These improvements, as well as others, have considerably increased Qwen 3 capacities compared to its predecessor, Qwen 2, explains Alibaba. None of the Qwen 3 models is the head and shoulders above recent high-end models like O3 and O4-Mini of Openai, but they are nevertheless strong artists.
On CodeForces, a platform for programming competitions, the largest Qwen 3 model-QWEN-3-235B-A22B-has just beaten O3-Mini of Openai and Gemini 2.5 Pro from Google. QWEN-3-235B-A22B is also the best of O3-Mini on the latest version of L’Aime, a difficult mathematical reference, and BFCL, a test to assess the capacity of a model to “reason” on problems.
But Qwen-3-235B-A22B is not accessible to the public-at least not yet.
The largest Qwen 3 public model, QWEN3-32B, is still competitive with a number of owners and open models, including the DEEPSEEK R1 from Chinese Lab. QWEN3-32B exceeds the O1 model of OPENAI on several tests, including the LiveCodebench coding.
Alibaba says that Qwen 3 “excels” in tool call capacities as well as the following instructions and the copy of specific data formats. In addition to download models, Qwen 3 is available from cloud suppliers, including AI and Hyperbolic fireworks.
Tuhin Srivastava, co-founder and CEO of the AI Cloud Baseten host, said that Qwen 3 is another point in the trend line of open models by following the rhythm of closed source systems such as Openai.
“The United States doubles the restriction of flea sales in China and purchases in China, but models like Qwen 3 which are at the cutting edge of technology and open […] will undoubtedly be used at the national level, “he told Techcrunch.” This reflects the reality that companies build their own tools [as well as] Buy on the shelf via closed fashion companies like Anthropic and Openai. »»
