Chinese tech giant Alibaba has launched Qwen3, a new set of AI models that it claims can match — and sometimes surpass — top systems from Google and OpenAI.
The Qwen3 lineup includes eight models, ranging from 0.6 billion to 235 billion parameters, and most of them are — or soon will be — openly available for download on platforms like Hugging Face and GitHub. In AI terms, more parameters usually indicate more advanced reasoning capabilities.
The introduction of models like Qwen3 from Chinese developers is raising the bar for American AI companies and prompting policymakers to tighten restrictions on China’s access to high-end chips used to train such systems.
“We release and open-weight Qwen3, our latest large language models, including 2 MoE models and 6 dense models, ranging from 0.6B to 235B. Our flagship model, Qwen3-235B-A22B, achieves competitive results in benchmark evaluations of coding, math, general…”
Qwen3 models are described by Alibaba as “hybrid” models, capable of adjusting their reasoning depth depending on the complexity of the request. For basic tasks, they can respond quickly, while more advanced reasoning takes longer. This approach helps improve accuracy, though it may introduce some latency.
“We have seamlessly integrated thinking and non-thinking modes, offering users the flexibility to control the thinking budget,” the Qwen team noted. “This design enables users to configure task-specific budgets with greater ease.”
Some models also use a Mixture of Experts (MoE) architecture, which boosts computational efficiency. MoE works by splitting a task into parts and assigning them to specialized “expert” models that handle each component more effectively.
Qwen3 supports 119 languages and was trained on over 36 trillion tokens — covering content like textbooks, Q&A pairs, code, and AI-generated material. This massive training set gives it a broad knowledge base and boosts its versatility.
Compared to its predecessor Qwen2, Alibaba says Qwen3 shows major improvements. While not dramatically better than top-tier models like OpenAI’s o3 or o4-mini, Qwen3 holds its own in performance.
For example, the top model, Qwen3-235B-A22B, outperforms OpenAI’s o3-mini and Google’s Gemini 2.5 Pro on coding benchmarks from Codeforces and math tests like AIME and BFCL. However, this flagship model isn’t publicly accessible — at least for now.
The most powerful publicly available model, Qwen3-32B, still ranks highly and surpasses several closed and open alternatives, including OpenAI’s o1 on benchmarks like LiveCodeBench.
Alibaba reports that Qwen3 also “excels” in tool-calling capabilities as well as following instructions and copying specific data formats. Besides being downloadable, the models are hosted by providers like Fireworks AI and Hyperbolic.
Tuhin Srivastava, CEO of AI hosting company Baseten, remarked that Qwen3 adds momentum to the trend of open models challenging closed systems.
“The U.S. is doubling down on restricting sales of chips to China and purchases from China, but models like Qwen 3 that are state-of-the-art and open […] will undoubtedly be used domestically,” he said. “It reflects the reality that businesses are both building their own tools [as well as] buying off the shelf via closed-model companies like Anthropic and OpenAI.”