Qwen3.8-27B
Alibaba (Qwen) · Multimodal model
Qwen3.8-27B is a mid-size open-weight Qwen model that reads text, images and video and is built for coding and agent tasks. It is released under Apache 2.0 and small enough to run on one high-end GPU.
Current Open weights
ReleasedAug 14, 2026
Context window262K tokens native (up to 1M with YaRN)
Input → outputtext, image, video → text
API price—
Consumer access—
Best for
Running a strong coding and computer-use agent locally on a single GPU under a permissive license.
Overview
Qwen3.8-27B is a mid-size open-weight Qwen model that reads text, images and video and is built for coding and agent tasks. It is released under Apache 2.0 and small enough to run on one high-end GPU.
Strengths and limits
- Apache 2.0 license
- Dense 27B model with image and video input
- High reported agent scores (SWE-bench Pro 61.7, OSWorld-Verified 84.3)
- Thinking mode on by default, which adds latency and tokens
- Hosted API version listed as 'coming soon' on the model card
Published benchmark scores
SWE-bench Pro61.7
Source: huggingface.co
Terminal Bench 2.1 (Terminus)73.0
Source: huggingface.co
GPQA Diamond89.2
Source: huggingface.co
OSWorld-Verified84.3
Source: huggingface.co
Scores are as published by the source; settings (reasoning effort, tools, attempts) differ between labs, so compare with care. How to read benchmarks →