Alibaba on Monday unveiled Qwen3, a new family of AI models that the company claims matches — and sometimes outperforms — the best models from OpenAI and Google.
Most of the Qwen3 models are available now or will soon be downloadable under an “open” license through Hugging Face and GitHub. The models vary widely in size, from 0.6 billion to 235 billion parameters. In AI, more parameters typically lead to stronger performance, as they enhance a model’s problem-solving abilities.
The emergence of China-developed models like Qwen3 has increased competitive pressure on American AI labs, including OpenAI. It has also prompted U.S. policymakers to tighten restrictions on Chinese companies’ access to the advanced chips needed for training such models.
Alibaba describes Qwen3 as a “hybrid” model family, capable of switching between quick responses and deeper reasoning. Reasoning allows the AI to self-check facts, much like OpenAI’s o3 model, but comes at the cost of higher response times.
“We have seamlessly integrated thinking and non-thinking modes, offering users the flexibility to control the thinking budget,” the Qwen team wrote in a blog post. “This design enables users to configure task-specific budgets with greater ease.”
Some Qwen3 models also use a “mixture of experts” (MoE) architecture, a design that improves computational efficiency. MoE divides tasks into subtasks and assigns them to smaller expert models, making the system faster and more flexible.
Qwen3 supports 119 languages and was trained on a massive dataset containing nearly 36 trillion tokens. Tokens are the basic units of text that AI models process, and 1 million tokens roughly equals 750,000 words. The training material included textbooks, question-and-answer sets, code snippets, AI-generated content, and more.
Thanks to these upgrades, Alibaba says Qwen3 shows major improvements over its predecessor, Qwen2. While the Qwen3 models don’t dominate top competitors like OpenAI’s o3 or Google’s Gemini 2.5 Pro across all benchmarks, they perform strongly in direct comparisons.
For example, on Codeforces, a platform for competitive programming, the largest Qwen3 model — Qwen-3-235B-A22B — slightly outperforms OpenAI’s o3-mini and Google’s Gemini 2.5 Pro. It also leads in benchmarks like AIME, a tough math test, and BFCL, which assesses reasoning capabilities.
However, Qwen-3-235B-A22B is not publicly available yet.
Among the models that are public, Qwen3-32B stands out. It competes favorably with leading models, including DeepSeek’s R1, another Chinese-developed system. Qwen3-32B also beats OpenAI’s o1 model in several tests, such as LiveCodeBench, a coding benchmark.
Alibaba says that Qwen3 shines in tool-calling, instruction-following, and replicating specific data formats. Beyond open downloads, the models are available through cloud services like Fireworks AI and Hyperbolic.
Tuhin Srivastava, co-founder and CEO of AI cloud platform Baseten, said Qwen3 is a strong sign of open models catching up with their closed-source rivals.
“The U.S. is doubling down on restricting sales of chips to China and purchases from China, but models like Qwen3 that are state-of-the-art and open … will undoubtedly be used domestically,” Srivastava told TechCrunch. “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.”
As the global AI race heats up, Alibaba’s Qwen3 series could reshape competition, challenging both technical benchmarks and political strategies.
