Alibaba unveils Qwen 3, a family of ‘hybrid’ AI reasoning models

Chinese Technical Company Alibaba Monday released Qwen 3, the family of AI models, claims the company’s matches, and in some cases goes to the best models available from Google and Openai.

Most of the models are available to download under “Open” license from the AI ​​Dev platform soon Hug face and Entrusted. They range from 0.6 billion parameters to 235 billion parameters. Settings are approximately a model of problem solving skills and models with more parameters generally perform better than less parameters.

The increase in the Chinese-formed model series, such as QWEN, increased pressure on American laboratories such as Openai to present more skilled AI technologies. They also led to the implementation of politicians to restrict the ability of China to acquire EU companies chips necessary Raising models.

According to Alibaba, Qwen 3 models are “hybrid” models in a sense that they can respond to time and “cause” or simpler desires through complex problems. Provisional models allow the facts to check the facts, similar to Openai models O3But with a higher delay value.

“Users offer comfort and thinking modes, offer comfort and thinking modes to control the budget budget, we offer flexibility to manage the budget of thinking,” said Qwen team a Blog Post. “This design allows users to configure task-specific budgets in an easier configuration.”

Qwen 3 models support 119 languages, says Alibaba and has been training on a data set of about 36 trillion Token. Tokens, the raw bits of information that has a model process; 1 million Token is about 750,000 words. Alibaba says that Qwen 3 textbooks, “question-answer pairs” code pieces, AI data, etc.

These progress, along with others, increased the performance of a large Qwen 3 compared to Alibaba, Gwen 2. Kodorkorces, a platform for programming competitions, Gwen-3-235B-A22B – Openai’s O3-Mini and Google’s Gemini and Google Twins defeated 2.5 Pro. QWEN-3-235B-A22B, AIME, a difficult math benchmark and BFCL, the latest version of a test to assess the “cause” ability about problems is the best of O3-mini.

However, QWEN-3-235B-A22B is not open – at least not yet.

The largest folk qwen 3 model, QWEN3-32B, China AI Lab Deepseek’s R1 is still competitive with a number of ownership and open AI models. Qwen3-32B exceeds Openai’s O1 model in several tests, including a precise benchmark, a precise benchmark.

Alibaba, Qwen 3 “Excel” says tool-challenge opportunities, as well as instructions and specific information formats. In addition to releasing models to download, Qwen 3 can be obtained from cloud providers, including fireworks AI and hyperbolic.

AI Cloud’s host market co-founder and CEO Tuhin Srivastava, Gwen 3, in the open 3 of the open models of PACE open models, said there is another point.

“The United States restricts the sale and purchases of chips in China, but the most up-to-date and open QWEN 3 models […] Surely it will be used locally, “he said. [as well as] Link off the shelf through indoor model companies such as anthropic and Openai. “