Qwen's Qwen3.8 collection on Hugging Face contains several downloadable model variants, including a 27-billion-parameter vision-language model and a much larger 2.4T-A95B mixture-of-experts model. The collection also lists FP8 versions of both.
The Qwen3.8-27B model supports text, image and video input. Its model card includes serving instructions for Transformers, vLLM and SGLang, and states a native context length of 262,144 tokens with configurable reasoning effort.
Qwen publishes a range of benchmark results for the 27B model, including coding-agent evaluations and visual tasks. In its own Terminal-Bench 2.1 setup it reports 73.0, compared with 63.4 for Qwen3.6-27B. Those figures depend on the listed benchmark conditions; they should not be read as guaranteed performance in a particular coding agent or deployment.
The Qwen3.8-27B listing identifies its license as Apache 2.0.