As a 27-billion-parameter dense member of the Qwen3.8 family, this model represents the most capable open-weights generation in the Qwen lineup to date. It inherits the hybrid-attention architecture from Qwen3.5, splitting 64 total layers into 16 layers of full gated attention and 48 layers of linear Gated DeltaNet attention with a constant recurrent state. A built-in multi-token prediction draft head accelerates inference, and a native 262,144-token context window supports long-horizon reasoning across professional and research workloads.

Qwen3.8-27B is a native vision-language model, structured as Qwen3_5ForConditionalGeneration with an explicit vision configuration, allowing it to process images and videos alongside text. It is positioned for substantial gains over Qwen3.5 and Qwen3.6 in coding, professional work, and autonomous agentic tasks, with stronger planning and end-to-end task completion. Thinking mode is on by default and can be tuned via reasoning effort levels, making the model a practical fit for developers building long-context assistants, code agents, and multimodal pipelines on open infrastructure.