As the AI race between the US and China heats up, Chinese tech giant Alibaba has expanded its open-source ambitions with the launch of four new small models under its Qwen 3.5 series. The lineup includes Qwen3.5-0.8B, 2B, 4B and 9B. These models are positioned as lightweight alternatives for developers who want multimodal capability and solid reasoning performance without running very large systems. And if the benchmarks are anything to go by, these models are giving tough competition to the nano and smaller offerings coming out of Silicon Valley heavyweights like OpenAI and Google.
All four models are built on the same Qwen 3.5 architecture and support native multimodal processing, meaning they can handle both text and images within a single model. Alibaba says the series includes architectural upgrades and scaled reinforcement learning. The company has released two versions of each model: a ‘base’ version for researchers who want to fine-tune from scratch, and an ‘instruct’ version for immediate deployment.
Tough competition to Gemini and ChatGPT
The most powerful model in this latest release is Qwen3.5-9B. According to benchmark results shared by Alibaba, the 9B model performs close to GPT-OSS-120B, despite being far smaller in size. In practical terms, Alibaba is claiming performance comparable to systems that power tools like OpenAI’s ChatGPT and Google’s Gemini in several reasoning and knowledge-based tests — but with a much smaller model.
In tests covering logical reasoning, maths problem-solving, multilingual knowledge and document understanding, the 9B version scores in the same range as much larger AI systems. It is also said to perform strongly in image-based reasoning tasks and document analysis benchmarks. In short, the model promises similar capability in a more compact form.
The second model, Qwen3.5-4B, sits just below it. Alibaba says this version comes close to the performance of its earlier 80B-parameter models. That makes it a practical option for developers who want solid reasoning and multimodal ability without running heavy infrastructure. It is positioned as a balanced choice — stronger than entry-level models, but lighter than full-scale systems.
At the smaller end, Qwen3.5-0.8B and 2B are designed for lightweight use. These versions are meant for phones, laptops and edge devices where computing power is limited. According to Alibaba, these models are faster and require less memory, though they are not as strong in complex reasoning as the 4B or 9B variants. Even so, they still support both text and image input, allowing them to handle basic multimodal tasks.
Alibaba has released the models publicly on Hugging Face and ModelScope with open weights available. Developers can download the checkpoints and run them locally using common AI frameworks.
Meanwhile, the launch of these small model series of Qwen is also drawing attention from the wider AI community. xAI boss Elon Musk reacted to the results on X, calling it “impressive intelligence density” in response to the benchmark comparisons shared for the latest Qwen 3.5 small series.
The latest launch from Alibaba adds to the growing competition around compact AI systems. Instead of focusing only on bigger and more powerful models, companies like Google, ChatGPT and others are now pushing smaller, more efficient versions such as Gemini Nano and lightweight ChatGPT variants. With Qwen 3.5 small models, Alibaba is targeting at developers who are building on-device AI tools, lightweight reasoning systems and multimodal applications without relying entirely on large cloud-based models.
