Alibaba's new generation model Qwen3.5 exposed

On February 9th, the latest PR (pull request for code merge) for Qwen3.5 was found on the open-source project page of HuggingFace, the world's largest AI open-source community, integrated into Transformers. Industry speculation suggests that Alibaba's new generation foundational model Qwen3.5 is about to be released, sparking heated discussions in the global AI open-source community. Some comments indicate that the "Crazy February" led by China's large models is about to begin. Related information reveals that Qwen3.5 adopts a brand new hybrid attention mechanism and is highly likely to be a VLM-type model capable of native visual understanding. Developers have further uncovered that Qwen3.5 may open-source at least a 2B dense model and a 35B-A3B MoE model. This latest development confirms previous reports by the U.S. media outlet The Information that Qwen3.5 will be open-sourced during the Spring Festival. Additionally, at the beginning of February, Tang Jie, Chief Scientist at Zhizhu, also mentioned on Weibo that many new models, including DeepSeek v4, Qwen3.5, and GLM-5, will be launched soon