Chinese tech company Alibaba on Monday Issued A family of AI model Qwen3, the company claims that matches claim and in some cases the best models available from Google and Openai have been improved.
Most models are – or soon – AI Dev will be available for download under the “Open” license from the platform Throat face And GithubThey range from 0.6 billion parameters to 235 billion parameters in size. Parameters correspond to the problem-coordination skills of a model broadly, and models with more parameters usually perform better than people with low parameters.
The rise of a China-origin model chain like Qwen has increased pressure on American laboratories such as OpenaiI to distribute more capable AI technologies. He has led policy makers, which is to implement restrictions aimed at limiting the capacity of Chinese AI companies to obtain the chips required to train the model.
Introduction Qwen3!
We release and open-weight Qwen3, our latest large language models, including 2 Mo models and 6 dense models, ranging from 0.6B to 235B. Our major models, Qwen3-235B-A22B, get competitive results in benchmark assessment of coding, mathematics, general … pic.twitter.com/jwzkjehwcc
– Qwen (@Alibaba_QWen) 28 April, 2025
According to Alibaba, the Qwen3 model is the “hybrid” model in the sense that they can take time and “cause” through complex problems or respond quickly to simple requests. The logic enables the model to make facts effectively, similar to models, but at the cost of high delay.
“We have basically integrated thinking and non-thinking modes, providing users flexibility to control the thinking budget,” one has written by QWEN team in one blog post“This design enables users to configure the working budget more easily.”
Some model experts (MOE) also adopt a mixture of architecture, which may be more computationally efficient to answer questions. Moe breaks tasks in subtask and assigned them to small, special “specialist” models.
The Qwen3 models support 119 languages, calling Alibaba, and trained on a dataset of around 36 trillion tokens. There are raw bits of tokens that process a model; 1 million tokens are equal to about 750,000 words. Alibaba says that Qwen3 was trained on combination with textbooks, “questions and answers pairs,” cod snipet, AI-related data, and more.
Alibaba says that these reforms greatly promoted the capabilities of Qwen3, compared to their predecessor, Qwen2, with others. Any of the QWEN3 model is the head and the recent models above the top-line above the head and the top-line of Openai, but they are still strong artists.
On Kodforce, a platform for programming competitions, the largest Qwen3 model-QWEN-3-235B-A22B-bus O3-Mini of Openai and Gemini 2.5 Pro of Google defeat the OPENAI’s O3-Min. Qwen-3-235B-A22B is also the best version of AIME, a challenging mathematics benchmark, and the latest version of BFCL is also the best O3-Mini, a test to assess a model’s ability to “cause” about problems.
But Qwen-3-235B-A22B is not publicly available-not at least yet.
The largest public QWEN3 model, Qwen3-32B, is still competitive with many ownership and open AI models, including R1 of Chinese AI Lab Deepsek. Qwen3-32B crosses the O1 model of Openai on several tests, including coding benchmark livecodebench.
Alibaba says that in tool-calling capabilities, Qwen3 follows “Excel” as well as following the instructions and mimics specific data formats. In addition to the model for download, Qwen3 is available from cloud providers, including fireworks AI and hyperbolic.
Tuhin Srivastava, co-founder and CEO of AI Cloud Host Besten, said that the open model with closed source systems such as Qwen3 OpeniI has another point in the line.
He said, “America is restricting the sale of chips to China and doubles purchasing from China, but models such as Quven 3 that are state -of -the -art and open … will undoubtedly be used domestically,” he said. “It shows reality that both businesses are buying their own devices (as well) from the shelf through closed-model companies such as anthropic and openi.”
