Qwen3.8-27B
27B dense, vision-capable, Apache 2.0 model reaching SWE-bench Pro 61.7 and OSWorld-Verified 84.3 with hour-scale video understanding.
A single-GPU-class Qwen3.8 with native image and video understanding, tunable reasoning effort and preserved thinking. Complements the 2.4T Max-class weights as the practical open option.
- Date
- Friday, 14 August 2026
- Lab
- Alibaba (Qwen)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench Pro / OSWorld-Verified | 61.7% / 84.3% Context 262,144 native, extensible to ~1M | company |
GitHub README date 2026-08-14 (HF repo created 2026-08-05); numbers (SWE-bench Pro 61.7, OSWorld-Verified 84.3) match the card's table. Model Studio lists the hosted qwen3.8-27b on 2026-08-17. Apache 2.0.
Sources
- huggingface.co/Qwen/Qwen3.8-27B
- github.com/QwenLM/Qwen3.8
- www.alibabacloud.com/help/en/model-studio/newly-released-models
This record was checked against its sources on 6 October 2026. How we check