On April 22, 2026, Alibaba Cloud's Tongyi Qianwen team made a groundbreaking announcement—the open-sourcing of Qwen3.6-27B, a robust, dense multimodal model boasting 27 billion parameters. This model marks a substantial leap forward in agent programming capabilities, comprehensively outperforming its forerunner, the open-source flagship model Qwen3.5-397B-A17B (a hybrid mixture of experts model with a staggering total of 397 billion parameters, of which 17 billion are actively utilized).

Qwen3.6-27B is inherently designed to support multimodality, seamlessly handling a blend of image, video, and text inputs. It is adept at tackling tasks such as visual reasoning, document comprehension, and visual question answering. In a series of programming benchmark assessments, including SWE-bench, Terminal-Bench 2.0, and SkillsBench, Qwen3.6-27B has showcased exceptional prowess.

As a model with a dense architecture, Qwen3.6-27B eliminates the necessity for MoE (Mixture of Experts) routing, thereby simplifying the engineering implementation process. Presently, the open-source weights of this model have been made available on the Hugging Face and ModelScope platforms, facilitating local deployment. Users can now enjoy online interactive experiences with Qwen3.6-27B on Qwen Studio, and the Alibaba Cloud BaiLian platform is set to offer API call support in the near future.