Wedoany.com Report on Mar 3rd, Alibaba's Qwen AI research team recently unveiled its newly developed Qwen3.5 small model series, including versions with parameter scales of 0.8B, 2B, 4B, and 9B. These models adopt a hybrid architecture design, combining Gated Delta Networks and Sparse Mixture of Experts technology, aiming to enhance efficiency and support native multimodal processing.

In multiple benchmark tests, the Qwen3.5-9B model performed excellently. In the GPQA Diamond benchmark, the model scored 81.7 points, surpassing the larger gpt-oss-120b model. In the MMMU-Pro visual reasoning test, it achieved a score of 70.1, leading ahead of Gemini 2.5 Flash-Lite. Furthermore, in the Video-MME video understanding evaluation, the 9B version scored 84.5, and the 4B version scored 83.5.

The developer community has responded positively to the Qwen3.5 small model series. Paul Couvert from Blueshell AI stated: "Qwen released 4 new models, the 4B version is almost as powerful as the previous 80B A3B. And the 9B model is as good as GPT OSS 120b, but 13x smaller!" He emphasized that these models can run on standard laptops, with the 0.8B and 2B versions suitable for mobile devices. Karan Kendre from Kargul Studio added: "These models can run locally on my M1 MacBook Air for free."

The model series has been released under the Apache 2.0 license on the Hugging Face and ModelScope platforms, permitting commercial use, modification, and distribution. Small AI models hold significant importance in current technological development; they support edge computing and can be applied in scenarios such as visual workflow automation, document parsing, and real-time analysis. For enterprise deployment, operational factors like memory requirements and debugging difficulty need consideration, and prioritizing verifiable tasks is essential to ensure output accuracy.