Tongyi Qianwen Qwen3.6-27B Open-Sourced: 27B Parameter Dense Model, Programming Capability Surpasses 15x Scale MoE Model

Alibaba Cloud's Tongyi Qianwen team recently announced the open-source release of Qwen3.6-27B, a dense multimodal model with 27 billion parameters—one of the most anticipated model specifications in the community. Before this, the team had launched versions like Qwen3.6-Plus; this 27B version keeps the dense architecture's advantages while greatly improving agent programming and multimodal reasoning abilities.

Official data shows Qwen3.6-27B performs exceptionally well in programming benchmarks. With just 27 billion parameters, it outperforms the Qwen3.5-397B-A17B MoE model (15 times its parameter size) across all major programming benchmarks. For instance, it scored 77.2 on SWE-bench Verified and 59.3 on Terminal-Bench 2.0, both higher than the previous generation.

The model natively supports multimodal tasks, handling mixed inputs of images, videos, and text. Its open-source weights are now available on Hugging Face and ModelScope platforms. Developers can download and deploy it, or interact via Qwen Studio or call it through Alibaba Cloud's Bailian platform API.