The Qwen 3.8 27B model has been released, featuring open weights and representing the latest advancement in the Qwen open-model family. Building on the Qwen3.5 architecture, Qwen3.8 offers significant improvements in coding, professional tasks, research, and long-horizon agentic operations. The Qwen3.8-27B variant is a compact, deployment-friendly dense model with native vision-language capabilities, understanding both images and videos, and designed for complex, multi-step tasks.

Key enhancements include improved autonomous planning, better handling of environment feedback for reliable task completion, and broader compatibility with development tools. It boasts a default thinking mode that can be adjusted, a context length of 262,144 natively expandable to 1,000,000 tokens, and native support for image and video understanding. The model weights are available in FP8-quantized format, compatible with various frameworks like Hugging Face Transformers.

For users prioritizing managed inference, the official Qwen API service via Qwen Cloud will soon offer a hosted version of Qwen3.8-27B with production-ready features. The release positions Qwen 3.8 as a leading local dense model, outperforming previous versions and competitors like Muse Glimmer-30B and Opus4.6 Max in various benchmarks.