Alibaba drops Qwen3.5 small model family from 0.8B to 9B

Running multimodal AI locally without setting your hardware on fire has always been a tough balancing act. Alibaba Qwen is tackling that constraint head-on.

The team just released the Qwen3.5 Small Model Series, rolling out four compact sizes: 0.8B, 2B, 4B, and 9B parameters. Along with the tuned versions, Alibaba also dropped the raw Base models on Hugging Face and ModelScope.

Big brains, tiny footprints

Despite their small footprint, these models share the same foundational upgrades as the broader Qwen3.5 family, including native multimodal processing, an improved architecture, and scaled reinforcement learning.

Each model has a distinct job. The 0.8B and 2B variants are built for sheer speed on resource-constrained edge devices. The 4B model is pitched as a capable multimodal backbone for lightweight agents, while the 9B model aims to close the performance gap with significantly larger systems.

Why it matters

For developers building local software or on-device gadgets, this lineup provides a versatile toolkit. You can deploy real multimodal reasoning on cheap hardware without latency spikes or hefty cloud bills.

We think releasing the Base models alongside the instruct variants is the right move here. It makes customized edge fine-tuning much easier, pushing the open-source battle line from cloud-hosted giants straight into local and on-device agent deployment.