Alibaba, a Chinese tech company, announced its launch on Monday Qwen3 is a family AI models that the company claims to be able to match and, in certain cases, even outperform models from Google and OpenAI.
The majority of the models will be available to download on AI dev platforms under an “open” licence. Hugging Face Github. They range from 0.6 billion to 235 billion. Models with more parameters perform better in general than those with less parameters.
As a result of the rise of China-originated models like Qwen, American labs like OpenAI are under more pressure to develop AI technologies that are more capable. These restrictions have also led policymakers in China to implement measures to limit the ability of Chinese AI firms to obtain the chips needed to train models.
Qwen3 is here!
Qwen3, the latest large language model, is released and open-weighted. It includes 2 MoE models as well as 6 dense models ranging from 0.6B up to 235B. Our flagship model, Qwen3-235BA22B, achieves excellent results in benchmark evaluations for coding, math and general… pic.twitter.com/JWZkJeHWhC
— Qwen (@Alibaba_Qwen) April 28, 2025
Alibaba says that the Qwen3 models can “reason” complex problems or respond quickly to simpler requests. The models can effectively fact-check their own answers, similar to OpenAI’s O3, but with a higher latency.
The Qwen team stated that “we have seamlessly integrated thinking modes and non-thinking mode, giving users the flexibility to control their thinking budget.” blog post. This design allows users to set up task-specific budgets more easily.
Other models adopt a mixture experts (MoE), which can be computationally efficient when answering queries. MoE divides tasks into smaller subtasks, which are then delegated to smaller “expert” models.
Alibaba said that the Qwen3 models supported 119 languages and were trained using a dataset of more than 36 trillion tokens. Tokens are raw data bits that a model processes. 1 million tokens is equal to approximately 750,000 words. The company said Qwen3 had been trained using a combination textbooks, “question and answer pairs,” code snippets as well as AI-generated data.
These and other improvements have greatly improved Qwen3 compared to Qwen2, Alibaba stated. The Qwen3 models are not as powerful as the latest models from OpenAI, such as the o3-mini and o4 mini. But they’re still very good.
In Codeforces, a platform used for programming contests and competitions, the Qwen3’s largest model, Qwen-3-235B A22B, just beat OpenAI’s o3 Mini and Google’s Gemini 2. 5 Pro. Qwen-3-235B A22B also beats o3 Mini on the latest versions of AIME and BFCL. These tests assess a model’s “reasoning” ability.
Qwen-3-235B A22B is not publicly available – at least, not yet.
Qwen3-32B is the largest public Qwen3 version. It is still competitive against a number proprietary and open AI models including DeepSeek R1 from Chinese AI lab DeepSeek. Qwen3 32B outperforms OpenAI o1 on several tests including the LiveCodeBench coding benchmark. Alibaba stated that Qwen3 excels in the ability to call tools, follow instructions and copy specific data formats. Qwen3 models are also available through cloud providers such as Fireworks AI and Hyperbolic.
Tuhin Srivastava is the co-founder and CEO at AI cloud host Baseten. He said Qwen3 was another example of how open models are keeping up with closed-source systems like OpenAI.
He told TechCrunch that the U.S. is increasing its efforts to restrict sales of chips and purchases from China. However, models like Qwen 3 which are open and state-of-the art[…] would undoubtedly be used in the United States. It reflects that businesses are building their own tools and buying off-the-shelf via closed model companies like Anthropic or OpenAI.
Kyle Wiggers, TechCrunch’s AI Editor. His writings have appeared in VentureBeat, Digital Trends and a variety of gadget blogs, including Android Police and Android Authority. He lives in Manhattan, with his music therapist partner.
View Bio
