Qwen releases Qwen3.8-27B, a dense 27B vision-language model under Apache 2.0
Qwen3.8-27B is a dense 27B model with a vision encoder, 262,144-token native context and switchable thinking, under Apache 2.0.
Qwen published Qwen3.8-27B, the compact dense member of its Qwen3.8 generation. The card calls Qwen3.8 the most capable generation in the Qwen open-model family to date. It is a causal language model with a vision encoder, so it takes images and video as well as text. The layout mixes Gated DeltaNet and gated attention layers, 64 layers in all.
Context is 262,144 tokens natively and extensible to 1,000,000. Thinking is on by default and can be switched off per request, and depth can be tuned with a reasoning_effort setting. The card says the weights work with Hugging Face Transformers, vLLM and SGLang. A hosted 1M-context version is described as coming soon, which is not the same as available.
The licence is Apache 2.0. The card says it builds on the Qwen3.5 architecture, and its main claims are better coding, research and long-horizon agent behaviour.
The card gives no release date; the date used is the Hugging Face repo creation date.
A permissive licence on a model people can actually run. The benchmark tables are the maker's own.
Sources
Editor's first pass (AI-assisted)
- Real 4.5: Ungated Apache 2.0 weights; the card lists Transformers, vLLM and SGLang support.
- Useful 4: Coding, agent tasks, images and video in one dense model that fits one machine more easily than the giants.
- Notable 4.5: Over 18,000 likes on Hugging Face, the most of any model reviewed here.
- Novel 3: Same architecture foundation as Qwen3.5, per the card; the gain is post-training.
Comments