Chinese high-tech company Alibaba released QWEN3 on Monday, a family of AI models it claims.
Most models can be downloaded under the “open” license on Face and Github, the AI Dev platform. The size ranges from 0.6 billion to 235 billion parameters. (Parameters are roughly compatible with the model’s problem-solving skills, and models with more parameters generally perform better than those with fewer parameters.)
The rise of Chinese-originated model series like Qwen has increased pressure on American labs such as Openai, providing more capable AI technology. They also led policymakers to implement restrictions aimed at limiting the ability of Chinese AI companies to obtain the chips needed to train models.
Introducing QWEN3!
Released QWEN3, the latest large-scale language model, which includes two MOE models and six dense models ranging from 0.6B to 235B and open weight QWEN3. The flagship model QWEN3-235B-A22B achieves competitive results with benchmark assessments on coding, mathematics, general…pic.twitter.com/jwzkjehwhc
– Qwen (@alibaba_qwen) April 28, 2025
According to Alibaba, the QWEN3 model is a “hybrid” model. It takes time to “infer” through complex problems or you can respond quickly to simpler requests. Inference allows models to effectively fact-check the model, similar to models such as OpenAI’s O3, at the expense of higher latency.
“We have a seamlessly integrated thinking and non-thinking mode. It gives users the flexibility to control their thinking budget,” the Qwen team wrote in a blog post. “This design makes it easier for users to configure task-specific budgets.”
Some models also employ a mixture of more computationally efficient expert (MOE) architectures to answer queries. Moe breaks down tasks into subtasks and delegates them to a smaller, specialized “expert” model.
The QWEN3 model supports 119 languages and was trained on a dataset of over 36 trillion tokens, Alibaba said. (Tokens are raw bits of data that the model processes. A million tokens equals about 750,000 words.) The company said QWEN3 was trained with a combination of textbooks, “question answer pairs,” code snippets, AI generated data, and more.
These improvements, along with others, have significantly improved the functionality of QWEN3 compared to its predecessor, QWEN2, Alibaba said. Both QWEN3 models don’t look more head and shoulder-like than the latest recent models like Openai’s O3 and O4-Mini, but they still have strong performance.
Programming contest platform CodeForces defeated the largest QWEN3 models, the O3-Mini and Google’s Gemini 2.5 Pro. QWen-3-235B-A22B bests O3-MINI with the latest version of AIME, a challenging mathematics benchmark, a test to assess the ability of a model on a problem.
However, Qwen-3-235B-A22B is not available, but at least it is not available yet.
The largest public QWEN3 model, the QWEN3-32B remains competitive with many unique and open AI models, including the R1 of the Chinese AI Lab Deepseek. The QWEN3-32B outperforms Openai’s O1 model in several tests, including the coding benchmark LiveCodebench.
Alibaba said QWEN3 is “good” with its ability to invoke tools, and “good” with the following instructions and copying certain data formats. In addition to the downloadable model, QWEN3 is available from cloud providers such as Fireworks AI and Hyperbola.
Tuhin Srivastava, co-founder and CEO of AI Cloud Host Baseten, said QWEN3 is another point in the trendline for open models that maintain pace with closed source systems such as Openai.
“The US has doubled the limit on sales of chips to China and purchases from China, but models like the cutting edge and open Qwen 3 are […] He told TechCrunch. “It reflects the reality that companies are building their own tools. [as well as] Buy shelves through closed model companies such as Anthropic and Openai. ”
