Alibaba reveals QWEN 3, a family of “hybrid” thinking models

Chinese technology company Ali Baba on Monday Absolute QWEN 3, a family of AI models that the company claims to match, and in some cases it surpasses the best available models from Google and Openai.

Most models – or will be soon – are available for download under a “open” license from the AI ​​Dev platform Embroidery and Jaytab. It ranges in size from 0.6 billion teachers to 235 billion teachers. Parameters are almost compatible with problem solving skills in the model, and models with parameters generally perform better than those that contain less parameters.

The height of the chain of models made from China, such as QWEN, has increased pressure on American laboratories such as Openai to provide more capable artificial intelligence techniques. He also led politics to implement restrictions aimed at reducing the ability of artificial intelligence companies to obtain Chips necessary To train models.

According to Ababa, QWEN 3 models are “hybrid” models, meaning that they may take time and “cause” through complex problems or answer simpler requests quickly. Thinking enables models to verify effectively facts, similar to models such as Openai’s O3But at the expense of the higher cumin.

QWEN wrote in A “We have smoothly thinking and thinkers, and we offer users to control the thinking budget.” Blog post. “This design enables users to create task budgets more easily.”

Alibaba says, QWEN 3 models support 119 languages, and have been trained in a collection of data of about 36 trillion symbol. The symbols are raw bits of the data that the model processes; One million symbols equivalent to about 750,000 words. Alibaba says that QWEN 3 was trained in a group of textbooks, “spouses of answering questions”, software excerpts, data created from artificial intelligence, and more.

These improvements, along with others, have greatly strengthened QWEN 3 compared to its predecessor, QWEN 2, that Ali Baba. On Codeforce, the programming platform for programming competition O3-Mini And Google’s Gemini 2.5 Pro. QWEN-3-235B-A22B is also the best O3-MINI on the latest version of AIME, which is a standard for difficult mathematics, and BFCL, a test to assess the model’s ability to “think” about problems.

But QWEN-3-235b-A22B is not available to the public-at least not yet.

The largest QWEN 3, QWEN3-32B model, is still able to compete with a number of royal and open artificial intelligence models, including Ai Lab Deepseek’s Chinese R1. QWEN3-32B exceeds Openai O1 Model in many tests, including LiveCodebeench coding standard.

Alibaba says QWEN 3 “Excels” in the possibilities of calling tools in addition to following instructions and copying specific data formats. In addition to downloading models, QWEN 3 of cloud service providers including AI and VyridLEC fireworks are available.

“QWEN 3 is another point in the direction line of open models in keeping up with closed systems such as Openai.

“The United States doubles the restrictions of chips on China and purchases from China, but models such as QWEN 3 will be used and open (…) undoubtedly locally,” he told Techcrunch in a statement. “It reflects the fact that companies are building their own tools (as well as) buying a shelf through closed -style companies such as Antarbur and Openai.”