Alibaba's Qwen Releases Qwen3.8-27B: Open Weights, Vision, 262K Context
Alibaba's Qwen team published Qwen3.8-27B in FP8 on Hugging Face — a compact dense model with native vision, 262,144-token context and tunable reasoning. It topped Hacker News within a day.
What happened
Alibaba's Qwen team published Qwen3.8-27B in FP8 on Hugging Face — a compact, deployment-friendly dense model from what Qwen calls its most capable open-model generation to date. The release hit the top of Hacker News on August 14, collecting more than 1,400 points in a day.
What is inside
Qwen3.8-27B is a native vision-language model: it understands images and videos, natively handles 262,144 tokens of context (extensible to one million), and offers flexible thinking control — thinking mode is on by default and reasoning depth can be tuned per request. The FP8 build uses fine-grained block-wise quantization and, per Qwen, performs nearly identically to the original weights while running on Transformers, vLLM and SGLang.
Why it matters
Built on the Qwen3.5 architecture, the new generation targets coding, professional work, research and long-horizon agentic tasks. A hosted version with a 1M-token context and built-in tools is coming to Qwen Cloud. For teams, the takeaway is practical: near-frontier agent workflows can now run on a self-hosted 27B model — without shipping data to a closed API.