Milestones
- Tongyi Qianwen unveiledApr 2023Complete.
- First open-weight Qwen-7BAug 2023Complete.
- Qwen2.5 familySep 2024Complete.
- Qwen3 hybrid-reasoning family29 Apr 2025Complete.
- Qwen3-Max announced at Apsara conference24 Sep 2025Complete.
- Qwen4 releaseNowCurrent milestone.
Most important updates
- 18d ago
- 3 Aug 2026
- Feb 2026
- 24 Sep 2025
- 29 Apr 2025
Upcoming
- Q4 2026Qwen-Image-3.1, which Alibaba said at Apsara will launch before the end of 2026 (next)
Current obstacles
- Access to advanced chipsUS export limits on NVIDIA chips push Alibaba toward its own and other Chinese chips for training.
Physics limits
- Running out of human-written textFrontier models already train on 15 to 40 trillion tokens. Estimates put the usable stock of public human text at a few hundred trillion, so data, not chips, starts to cap plain scaling.
- Shrinking numbers loses informationRunning on smaller hardware means storing weights in 4 to 8 bits instead of 16. Below about 4 bits accuracy drops quickly, so model size sets a floor on memory needed.
- Returns shrink as a power lawError falls only as a small power of training compute: each fixed step of improvement needs roughly 10x more compute, energy and money than the one before.
How it works
Every size, mostly open
Qwen ships dense models from under 1B up to 235B-parameter mixtures of experts, most under Apache 2.0, so it is a common base for others.
Thinking on demand
Qwen3 models can answer instantly or reason at length within the same model, switched by the user or by a thinking budget.
36 trillion tokens
Qwen3 was pretrained on about 36 trillion tokens covering 119 languages and dialects, heavy in code and maths.
Largest models via cloud
The biggest models, such as Qwen3-Max with over a trillion parameters, are served through Alibaba Cloud rather than released as weights.
Update log
Tue 22 Sep
- Major: PressResearch
Mon 3 Aug
- Minor: OtherSoftware
Feb 2026
- Minor: OtherSoftware
Wed 24 Sep 2025
- Minor: PressSoftware
Tue 29 Apr 2025
- Major: BlogSoftware
Tue 28 Jan 2025
- Minor: BlogSoftware
Thu 19 Sep 2024
- Minor: BlogSoftware
About Alibaba Qwen
Alibaba Qwen
Qwen open-weight models, Qwen app
Qwen is Alibaba Cloud's AI model family and one of the most used open-weight lines, with hundreds of thousands of spin-off models online. Sizes range from under 1 billion to over 1 trillion parameters.
- Qwen3 (April 2025) can answer fast or think step by step, and is open under a permissive licence.
- Alibaba committed RMB 380B (about $53B) over three years to cloud and AI in February 2025.
- Qwen3-Max, the closed flagship, has over a trillion parameters.
- Founded
- 20233 yrs
- Headquarters
- China
- Status
- Subsidiaryof Alibaba Group
- Valuation
- Alibaba Group-owned
- Works in
- AIOpen-weight models
- Coverage
- 1 program · 8 updateslatest 18d agochecked 25 Sep