Recently, Alibaba Qwen announced that the Qwen3-VL family has welcomed new members, adding two dense model sizes: 2B and 32B. This expansion covers visual language understanding scenarios from lightweight to high-performance, allowing developers to run these models on devices such as smartphones, greatly facilitating development and application.

image.png

The two model versions introduced this time each have their own features. First is the Instruct model, which has advantages in fast response and stable execution, making it especially suitable for dialogue systems and tool calls. The second is the Thinking model, which excels in long-chain reasoning and complex visual understanding, possessing the ability to "think while looking at images," and can handle more challenging tasks.

According to official release information, Qwen3-VL-32B outperforms some competitors in the market, such as GPT-5mini and Claude4Sonnet, in multiple fields. It can match models with up to 235B parameters using only 32B parameters, and even achieved excellent results on OSWorld. At the same time, Qwen3-VL-2B, with its compact size, can achieve surprising performance on extreme edge devices, making it suitable for developers to experiment and deploy.

For developers interested, Alibaba Tongyi also provides an experience link, making it convenient for users to try these new models on ModelScope and Hugging Face. The launch of these models not only expands Alibaba Tongyi's product line in the field of artificial intelligence, but also provides more possibilities for applications in visual language understanding.

Key Points:   

🌟 New Models: The Alibaba Qwen3-VL family has added two dense model sizes: 2B and 32B.  

📱 Device Compatibility: New models can run on devices such as smartphones, making them convenient for developers to use.  

🏆 Excellent Performance: Qwen3-VL-32B performs better than many competitors in multiple fields.