The Qwen team has announced Qwen3.8 as the highest-performing generation within their open-weight model family.
Following the widespread adoption of the Qwen3.5 series, this release achieves significant performance improvements in coding, specialized professional tasks, research, and long-term agent tasks.
Qwen3.8-27B is a compact, easy-to-deploy dense model. It is designed as a native vision-language model capable of understanding images and videos.
It is engineered to complete complex, multi-step tasks with higher reliability.
By default, a thinking mode is enabled, generating a reasoning process before the final answer.
The balance between reasoning depth and cost can be adjusted via the reasoning_effort parameter. Three levels are supported: xhigh, medium, and low.
In multi-turn agent tasks, lowering the reasoning effort results in faster responses per turn. However, this may lead to insufficient analysis, failures, or increased retries, potentially increasing overall latency and token consumption.
SGLang, vLLM, and TokenSpeed are recommended as inference frameworks. A hosted version on Qwen Cloud is also expected to be available soon.
The team stated that a 1M context length and official built-in tools will be available by default.
Source: Qwen3.8 27B (HN 65pt, 13 comments) (HN Search (backfill), 2026-08-14)