Qwen 3.5 Small Series

The Qwen 3.5 Small Model Series is Alibaba's answer to the edge AI era — a family of four open-weight, natively multimodal models built to run on consumer hardware without sacrificing modern reasoning capabilities.

**The model lineup**: - **Qwen3.5-0.8B** — Designed for smartphones and IoT hardware. Optimized for low VRAM and high throughput; suitable for on-device assistants, rapid prototyping, and fine-tuning experiments. - **Qwen3.5-2B** — Edge-tier. Balances latency and quality for embedded agents and lightweight edge deployments. - **Qwen3.5-4B** — The multimodal bridge. Processes text, images, and video natively, making it ideal for UI navigation agents, document analysis, and local copilots without the overhead of adapter-based approaches. - **Qwen3.5-9B** — The compact production flagship. Closes much of the gap with 30B+ models through advanced training; beats OpenAI's gpt-oss-120B on several benchmarks while running on a standard consumer GPU.

**What sets them apart**: - **Native multimodality** — Unlike earlier small models that bolted on vision encoders, Qwen3.5-4B and above process text, images, and video from the same weights — no adapters, no pipeline complexity. - **RL-trained judgment** — All four models are trained with reinforcement learning across massive agent environments, giving them human-aligned judgment for multi-step objectives like UI navigation, code generation, and agentic task execution. - **Truly open** — Available under Apache 2.0 on Hugging Face and ModelScope with both Instruct and Base versions.

The Qwen3.5 Small series democratizes agentic AI for hardware-constrained deployments — whether that's a smartphone, a Raspberry Pi cluster, or a developer's laptop.