Alibaba just launched Qwen3, a new series of open-source AI models that experts say could rival top systems from Google and OpenAI.
It’s a bold move in China’s growing push to dominate open-source AI.
Qwen3 isn’t just another large language model. It’s a hybrid reasoning system that blends rapid-response “non-thinking” with slower, deliberate “thinking” modes, giving developers control over performance vs. speed.
“We have seamlessly integrated thinking and non-thinking modes,” Alibaba said.
Models range from 0.6B to 235B parameters, and many are now free to download on Visage étreint, GitHub, and Alibaba Cloud. The largest version, Qwen-3-235B-A22B, beat OpenAI’s o3-mini et Google's Gemini 2.5 Pro on key coding and reasoning benchmarks, though it’s not yet publicly available.
Inside the AI
- Trained on 36T tokens
- Supports 119 languages
- Uses “mixture of experts” (MoE) architecture for efficiency
- Built for everything from mobile AI to cloud deployment
Photo by Solen Feyissa sur Unsplash
The race for dominance
Qwen3 shows that China’s AI labs are rapidly closing the gap with U.S. giants and doing it open source. It’s the latest escalation in the U.S.-China AI rivalry, despite American chip restrictions aimed at slowing Chinese progress.
“This is a significant breakthrough… despite mounting pressure from tightened U.S. export controls,” said analyst Ray Wang.
The rise
Alibaba says the Qwen series has been downloaded over 300M times, with 100K+ spinoff models on Hugging Face. With DeepSeek’s R2 on the horizon and Baidu pivoting toward open-source too, China’s AI momentum is hard to ignore.
Qwen3 proves that cutting-edge AI isn’t just coming from Silicon Valley anymore, and the open-source frontier may now be led by China.
