Qwen3.8-27B is the compact dense member of the Qwen3.8 family, the most capable generation in the Qwen open-model line to date, designed to carry complex, multi-step tasks through to completion with greater reliability. It is built on the architectural foundation of Qwen3.5, sharing the same hybrid-attention backbone used by the larger flagship: across 64 layers, 16 run full gated attention on a four-layer interval while the remaining 48 use linear attention via Gated DeltaNet with a constant recurrent state. This hybrid design supports long-context inference and sustained reasoning, and the model ships as a native vision-language model that understands images and videos alongside text.

Practically, Qwen3.8-27B is positioned for coding, professional work, research, and long-horizon agentic workloads, with stronger autonomous planning, improved handling of environment feedback, and broader compatibility with popular development harnesses. It includes a built-in Multi-Token Prediction draft head for faster inference and a native 262,144-token context window, with extensions to the cataloged API limit in hosted deployments. Weights and configurations are released in the Hugging Face Transformers format and run on Transformers, vLLM, SGLang, and TokenSpeed, while thinking mode is on by default with tunable reasoning depth. The combination of dense 27-billion-parameter scale, hybrid attention, and open weights makes it well suited for teams that want a deployable, full-featured Qwen3.8 model without operating the larger mixture-of-experts variant.