Alibaba just released the Qwen3.5 Small Model Series, and the numbers are hard to ignore. The flagship of this new lineup, Qwen3.5-9B, outperforms OpenAI’s gpt-oss-120B on key benchmarks — despite being 13.5 times smaller. It runs on standard laptops.
This isn’t a research prototype or a demo. The weights are available on Hugging Face and ModelScope right now, under Apache 2.0 licenses, ready for enterprise use.
The Models
The series includes four variants:
- Qwen3.5-0.8B and 2B — optimized for edge devices and battery-constrained deployment
- Qwen3.5-4B — multimodal base model with a 262,144 token context window
- Qwen3.5-9B — the standout, a compact reasoning model that beats models over 100x its size
The 9B model achieves an 81.7 on GPQA Diamond (graduate-level reasoning), surpassing gpt-oss-120B’s 80.1. On MMMU-Pro visual reasoning, it scores 70.1 against Gemini 2.5 Flash-Lite’s 59.7. In video understanding (Video-MME), it hits 84.5. These aren’t marginal wins — they’re decisive advantages across multimodal, reasoning, and knowledge tasks.
Why It Works
Alibaba moved beyond standard Transformer architecture. The Qwen3.5 Small Series uses an Efficient Hybrid Architecture combining Gated Delta Networks (a form of linear attention) with sparse Mixture-of-Experts. This addresses the “memory wall” that typically limits small models — higher throughput, lower latency, and significantly reduced memory requirements during inference.
The models are also natively multimodal. Previous generations bolted vision encoders onto text models. Qwen3.5 was trained using early fusion on multimodal tokens, meaning the 4B and 9B models can read UI elements, count objects in video, and understand visual context at a level that previously required models ten times their size.
Why This Matters
For years, the AI industry has been locked in an arms race — bigger models, more parameters, more compute. Qwen3.5-Small flips that narrative. “More intelligence, less compute” isn’t a slogan — it’s the benchmark results.
This has real implications:
Local-first AI becomes viable. Teams no longer need GPU clusters to run competitive models. A laptop can now handle tasks that six months ago required cloud infrastructure.
Edge deployment shifts from theory to practice. The 0.8B and 2B models target exactly the use case where battery life and device constraints matter — smartphones, embedded systems, on-premise servers without GPU racks.
The open-source ecosystem keeps gaining ground. Following DeepSeek’s momentum earlier this year, Qwen3.5-Small further demonstrates that open-weight models can match or exceed proprietary alternatives at a fraction of the cost and compute. Our open source AI models guide tracks how the leading open-weight options compare across categories, including which small models are worth running locally.
The Bigger Picture
We wrote recently about AI agents rewriting the automation playbook — the shift from models that answer questions to models that complete tasks. Qwen3.5-Small accelerates that shift. When a model this small can reason through graduate-level problems, understand video content, and operate with a 262K token context window, the barrier to deploying capable agents drops significantly.
The model weights are live on Hugging Face. If you’ve been waiting for the right moment to start running AI locally, this is it.