Qwen3.8-27B brings the Qwen3.8 generation to a dense, deployment-friendly scale. It accepts text, images, and video and is designed for coding, professional work, research, and long-horizon agentic tasks.

Model overview

• Dense 27B model: a causal language model with a vision encoder, built on the Qwen3.5 architecture.
• 262K native context: supports up to 262,144 tokens and can be extended toward 1M tokens with RoPE scaling.
• Flexible reasoning: thinking is enabled by default and can be disabled or adjusted with xhigh, medium, and low reasoning-effort levels.
• Preserved thinking: reasoning context from earlier messages is retained by default for continuity across multi-step work.
• Multimodal understanding: supports images and videos, including documents, STEM diagrams, and long-form video.

Qwen-reported benchmark results

Qwen reports substantial gains over Qwen3.6-27B across coding, agentic work, and multimodal tasks:

Benchmark	Qwen3.8-27B	Qwen3.6-27B
Terminal Bench 2.1	73.0	63.4
SWE-bench Pro	61.7	53.5
DeepSWE 1.1	42.2	13.3
QwenSWEBench	79.0	49.3
CoWorkBench	70.7	61.0
LiveCodeBench v6	90.3	83.9
OSWorld-Verified	84.3	63.9
WebArena-Verified	64.8	48.8
SWE-MM	38.6	25.7

These results are reported by Qwen. SWE-bench Pro and DeepSWE 1.1 use the Claude Code harness with temperature=1.0, top_p=0.95, and a 256K context window. QwenSWEBench and CoWorkBench are in-house evaluations. See the official Qwen3.8-27B model card for the complete benchmark tables and methodology.

License

Qwen3.8-27B is available under the Apache 2.0 License.