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

Chinese Technology Company Alibaba on Monday released Qwen 3, the AI ​​model family that the company claims to match, and in some cases it exceeds the best models available from Google and Openi.

Most of the models are – or will soon be – available to download under an “open” license with AI DeV platform Hug your face and Girlub. They move in size of 0.6 billion parameters up to 235 billion parameters. Parameters are approximately corresponding to the skills of solving models’ problems, and multi -parameter models generally seem better than those with fewer parameters.

The rise of the series of models originating in China, such as Qwen, has increased pressure on US laboratories such as OPENAI for delivery of more capable AI technologies. They also cited the policy creators to implement limits aimed at restricting the ability of Chinese AI companies to get chips needed to train the model.

According to Alibaba, Qwen 3 models are “hybrid” models in the sense that they can take time and “reason” through complex problems or respond quickly to simpler requirements. The explanation allows models to effectively check the facts, similar to models such as Openai’s O3, but at the cost of greater latency.

“We have an imperceptibly integrated ways of thinking and not thinking, offering users flexibility to think of budget for thinking,” the Qwen team wrote in a blog post.

Qwen 3 models support 119 languages, Alibaba says, and are trained at a set of data of nearly 36 token trillion. Tokens are the raw bits of the data that the model processes; 1 million tokens are equivalent to about 750 000 words. Alibaba says Qwen 3 is dressed at a combination of textbooks, “couples to answer questions,” clips of the code and more.

These improvements, along with others, are greatly enhanced by the Qwen 3 performance compared to Qwen 2 predecessor, Alibaba says. At Codeforces, Programming Platform, the largest model Qwen 3-Qwen-3-235b-A22B-Plims Openi’s O3-Dom. The QWEN-3-235B-A22B is also O3-Di on the latest version of AIME, a challenging mathematical measure and a BFCL, a test to assess the model’s ability to “distinguish” about problems.

But Qwen-3-235b-A22b is not publicly available-it is not yet.

The largest public Qwen 3 model, Qwen3-32B, is still competitive with numerous ownership and open AI models, including R1 of Chinese AI LAB Deepseek. The QWEN3-32B on several tests surpasses the OPENAI O1 model, including the reference value of accuracy called Livebench.

Alibaba says Qwen 3 “Excels” in the options of calling to the tool, as well as the following instructions and copying of specific data format. In addition to publishing the download model, the Qwen 3 is available from cloud service provider, including fireworks AI and Hyperbolic.

Tuhin skewers, co -founder and executive director AI Cloud Host Basen, said Qwen 3 is another point in the trendy line of open models that keep up with the closed code systems such as Openai’s.

“Now they double on to limit the sales of chips on China and buy from China, but models like Qwen 3 that are top -notch and open […] It will undoubtedly be used in the country, “he told Techcrunch in a statement.” This reflects the reality that companies both build their own tools [as well as] buy the shelf through companies of closed models such as Anthropic and Openi. “