Best for researchers and developers prototyping or fine-tuning on modest hardware, with free weights from under a billion parameters.
Kenning's take
What it does
The smaller Qwen3.5 models, at 0.8B, 2B, 4B and 9B parameters, were released on Hugging Face at the end of February 2026 alongside larger Qwen3.5 models. They share the series' unified vision-language design and hybrid architecture. Qwen says the smallest sizes suit prototyping, fine-tuning and research.
- 0.8B, 2B, 4B and 9B sizes
- Vision and language in one model
- Hybrid Gated DeltaNet and MoE design
- Base versions for fine-tuning
- Apache 2.0 open weights
Who it suits. Developers who want small open models to run on modest hardware or fine-tune.
Key facts
- Category
- AI models
- Part of
- Qwen
- Made by
- Alibaba (Qwen team)
- Based in
- China
- Released
- 27 February 2026
- How to get it
- Free download from Hugging Face; Qwen Chat for the family
- Works on
- self-hosted, Hugging Face
- In the UK
- Yes
- Free plan
- Yes, open weights
FAQ
What sizes are the small Qwen3.5 models?
0.8B, 2B, 4B and 9B parameters.
Are Qwen3.5 small models free?
Yes. The weights are open under Apache 2.0.
Facts from the maker's own pages and the sources listed. Checked 2026-10-11.
Similar tools
MAI-UI
Alibaba Tongyi research models that operate smartphone apps by looking at the screen and tapping
Qwen-Image-2.0
Alibaba's 7-billion-parameter model that both creates and edits images, with strong text rendering at 2K
Qwen-Image-Layered
Open Alibaba model that splits an image into separate transparent layers so each part can be edited alone
Qwen-Image-3.0
Alibaba's hosted image model built for text-heavy layouts, taking prompts up to 4,500 tokens