Alibaba / Qwen · 2026-03-02 · notable
Qwen 3.5 Small — on-device multimodal models
Four compact multimodal models (0.8B, 2B, 4B, 9B) under Apache 2.0 with 262K context — the first family where a 0.8B model can process video and a 9B beats previous-gen 30B models.
Four natively multimodal models from 0.8B to 9B parameters — small enough for phones, capable enough to beat last-gen 30B models.
Key specs
| License | Apache 2.0 |
|---|---|
| Context window | 262K tokens |
| Variants | 0.8B, 2B, 4B, 9B |
What is it?
The Qwen 3.5 Small Model Series is a set of four compact models released March 2, 2026, completing the Qwen 3.5 family rollout. The four sizes (0.8B, 2B, 4B, 9B) are all natively multimodal with 262K token context and Apache 2.0 licensing. This is the first model family where a 0.8B model can process video, a 4B model can act as a lightweight multimodal agent, and a 9B model beats previous-generation 30B models across the board.
How does it work?
All four models share the Qwen 3.5 architecture with native multimodal support, scaled RL training, and improved architecture design. The 0.8B and 2B are designed for edge devices and phones, the 4B serves as a base for lightweight agents, and the 9B is the power-efficient option for single-GPU inference. GGUF quantized variants are available on Hugging Face for all sizes.
Why does it matter?
On-device AI needs models that are both small enough to run locally and capable enough to be useful. These four models cover the full range from wearables to laptops, and the fact that even the 0.8B handles video means developers can build multimodal features without sending data to a cloud API.
Who is it for?
Mobile and edge developers, on-device AI teams, anyone building offline-capable AI features.
Try it
ollama run qwen3.5:9b