Read our How to Run Qwen3.8-27B Guide! This GGUF uses Unsloth Dynamic V3.0 (preview) for SOTA quantization performance. Developer Role Support enables Qwen3.8 to work in agentic tools like Codex and more.

Built on Qwen3.5's architectural foundation, Qwen3.8 delivers substantial gains across coding, professional work, research, and long-horizon agentic tasks. Qwen3.8-27B is a native vision-language model that understands images and videos, with flexible thinking control designed for complex multi-step task completion.

Key features include comprehensive coding and professional work improvements, stronger autonomous planning, broader compatibility with development tools, and native image/video understanding. The model supports context lengths up to 262,144 tokens natively and 1,000,000 with RoPE scaling.

Recommended settings include temperature=1.0 with top_p=0.95 for thinking mode, and temperature=0.7 with top_p=0.80 for instruct mode. For agentic tasks, allocate 262,144 tokens for reasoning and 131,072 for final response.