Chinese technology company Alibaba on Monday unveiled its Qwen3 series of artificial intelligence models , which – according to the company – can compete with or even surpass, in some cases, the top models from Google and OpenAI .
See also: Alibaba to invest $52 billion in AI and cloud computing
Most of these models are – or will soon be – available for download under an “open” license on the AI development platforms Hugging Face and GitHub. They range in size from 0.6 billion to 235 billion parameters. (Parameters are related to a model’s problem-solving ability; generally, the more they have, the better the performance.)
The rise of Chinese-made AI models like Qwen has increased pressure on U.S. labs like OpenAI to develop more powerful AI technologies, and has prompted policymakers to enact restrictions aimed at limiting Chinese AI companies’ access to specialized chips for training models.
According to Alibaba, Qwen3 models are “hybrid” – they can take their time to process complex problems or respond instantly to simple requests. The “reasoning” capability allows the models to check the accuracy of their answers, similar to OpenAI’s o3, but this comes with increased latency in response.
Some of the models also adopt a “mixture of experts” (MoE) architecture, which can be more computationally efficient for answering questions. MoE breaks down tasks into subtasks and delegates them to smaller, specialized “expert” models.
See also: China: Apple and Alibaba teaming up to bring AI to iPhone users?
According to Alibaba, the Qwen3 models support 119 languages and were trained on a dataset of nearly 36 trillion tokens. (Tokens are the basic units of data a model processes; 1 million tokens corresponds to about 750,000 words.) The training included textbooks, question-answer pairs, code snippets, AI-generated data, and more.
These improvements, combined with others, have significantly enhanced Qwen3's capabilities compared to its predecessor, Qwen2, according to Alibaba. While none of the Qwen3 models appear to clearly outperform current top-of-the-line models, such as OpenAI's o3 and o4-mini, they are nonetheless particularly strong options.
On the Codeforces programming competition platform , the largest model in the Qwen3 series — the Qwen-3-235B-A22B — manages to slightly outperform OpenAI’s o3-mini and Google’s Gemini 2.5 Pro. The Qwen-3-235B-A22B also outperforms the o3-mini on the latest version of the AIME, a demanding math benchmark, as well as the BFCL , a test that measures a model’s ability to “reason” about problems.
However, Qwen-3-235B-A22B is not available to the public — at least for now.
See also: Alibaba Cloud vulnerability allows data upload
An interesting point related to the above is that Alibaba’s strategy to rely on hybrid architectures such as Mixture of Experts (MoE) and train its models on huge amounts of data (36 trillion tokens) shows a clear intention to compete directly with Western giants such as OpenAI and Google. The fact that Qwen-3-235B-A22B outperforms their models in specific benchmarks — especially in areas that require a high level of logical thinking and mathematical ability — shows that China is not just aiming to follow developments in artificial intelligence, but to lead in critical areas.
Source: techcrunch
