Qwen · Model release
Qwen3.8-27B: Apache 2.0 dense VLM for local work
Qwen3.8-27B is a 27B dense vision-language checkpoint under Apache 2.0 — a practical local path that is not the hosted Qwen3.8-Max API.
What changed
- Hugging Face published Qwen/Qwen3.8-27B as an Apache 2.0 dense VLM with native 262,144-token context and image/video input.
- Thinking is on by default and can be disabled; reasoning_effort is xhigh / medium / low.
- The current QwenCloud managed card and Model Studio pricing list a hosted qwen3.8-27b SKU with 1M context; the older checkpoint card's coming-soon text is not current hosted availability proof. Native weights remain a separate 262,144-context deployment.
Best fit
- Local multimodal agents
- Apache 2.0 self-hosting
- Smaller-than-Max open deployments
What to be careful about
- Vendor SWE-bench Pro and DeepSWE figures used a Claude Code harness — they are not swebench.com or Datacurve public-board scores.
- Do not merge this checkpoint with hosted Qwen3.8-Max.
Dataset snapshot: Qwen 3.8 27B is also represented in the checked comparison dataset. Prices and benchmark figures carry their own field dates.