Qwen3.8-27B

Qwen3.8-27B is a 27-billion-parameter dense vision-language model from Alibaba's Qwen team, released in August 2026 under Apache 2.0. It builds on the Qwen3.5 architectural foundation with a hybrid design: 64 layers interleaving Gated DeltaNet linear-attention blocks with gated attention, trained with multi-token prediction. It natively understands images and video, including hour-scale video, and targets coding, professional work, research, and long-horizon agentic tasks. Thinking mode is on by default with per-request control, reasoning depth tunable via reasoning_effort, and reasoning context retained across messages via preserve_thinking. The context window is 262,144 tokens natively and extensible to 1,000,000.

The model card's capability claims (gains across coding, professional work, research, and agentic reliability) are provider-stated and have not been independently reproduced for this entry; what is independently verified is the checkpoint itself, the Apache 2.0 license, the architecture summary, and compatibility with Transformers, vLLM, SGLang, and TokenSpeed. A hosted version with 1M context by default and official built-in tools was announced for Qwen Cloud as a managed alternative to self-hosting.