Qwen3
Era 5 · Generative AI, reasoning, and agents (2022–2026) › Frontier models
- Lineage
- Neurons and networks
- Kind
- Implementation
Who
Alibaba
What happened
Alibaba releases Qwen3, a family of eight open models, from 600 million to 235 billion parameters, that combine a reasoning mode and a fast-response mode in the same model. They are trained on 36 trillion tokens and handle 119 languages and dialects.
Why it mattered
Open hybrid-reasoning models in every size, from the phone to the data center. The Qwen family becomes the most widely used base for derivative models.
The full story
After DeepSeek-R1, in January 2025, competition among Chinese labs sped up. Alibaba, China's largest e-commerce company, had been releasing its Qwen models with open weights since 2023.
On April 29, 2025, it presented Qwen3, its third generation. There were eight models: six dense ones, from 600 million to 32 billion parameters, and two mixture-of-experts models. The largest had 235 billion parameters and used 22 billion per token. All were released under the Apache 2.0 license.
What was new was hybrid reasoning. Like Claude 3.7 Sonnet a few months earlier, each model could think step by step on hard problems or answer right away on simple ones, and the user could set how much thinking time it got. It was trained on 36 trillion tokens, twice as many as its predecessor, and went from 29 to 119 languages and dialects. It also supported the Model Context Protocol for connecting to tools.
The scale of its use was the other headline. According to Alibaba, the Qwen family had already passed 300 million downloads, and developers had created more than 100,000 derivative models on the Hugging Face platform. Chinese open models, and not just DeepSeek, had become the foundation of much of the world's open ecosystem.
Builds on
Model Context Protocol (November 25, 2024)
Documented direct influence
Qwen3 supports the Model Context Protocol natively, and ChatGPT agent connects to outside apps through connectors
DeepSeek-R1 (January 20, 2025)
Inherited, with no proven direct link
After R1, other Chinese labs release open frontier models under permissive licenses
Threads that pass through here
- China and the chips (step 4 of 6)
Sources
Original source
- Qwen3 Technical Report (arXiv:2505.09388) (opens in a new tab). Equipo Qwen, 2025
Scholarship and context
How to cite
“Qwen3.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/qwen3
This entry was last reviewed on November 13, 2026.