On the evening of March 2nd, Alibaba made an official announcement, revealing that it had open-sourced four compact models within the Qwen3.5 series: Qwen3.5-0.8B, 2B, 4B, and 9B. Leveraging native multimodal training techniques and the latest architectural advancements, these models are designed to cater to a wide array of scenarios, spanning from environments with extremely limited resources to high-performance, lightweight applications. Specifically, the 0.8B and 2B models are tailored for seamless deployment on mobile devices, the 4B model emerges as the ideal choice for lightweight agents, and the 9B model stands out as a cost-effective solution for server-side inference. At present, all these models are readily available as open-source on the ModelScope community and Hugging Face platforms. Elon Musk even commented on their remarkable intelligence on the X platform, expressing his admiration.