Qwen 3.8 27B

Qwen's native vision-language dense model, building on the previous 27B generation with key improvements in coding and office productivity. It is a dense 27B model with a built-in vision encoder giving native image and video understanding alongside text, and it is the compact, deployment-friendly member of the 3.8 family. Thinking is on by default and can be disabled per request, with reasoning preserved across conversation history.

Alibaba shipped this one quietly. There is no launch post for the 27B: it was announced on 14 August 2026 through the lab's own Qwen3.8 repository and its model card, two days after the Max-class open weights, as the compact member of the same generation. Everything below therefore comes from the lab's own release notes and model card rather than from a blog write-up, and there is no marketing framing to discount.

It is a dense model, and that is the point of it. All 27 billion parameters run on every token, which costs more per token than a sparse model of similar size but makes it far simpler to deploy - one machine, predictable memory, no expert routing to tune. Alibaba describes 64 layers in a hybrid layout that runs three Gated DeltaNet blocks for every one full-attention block, with multi-token prediction trained in, native context of 262,144 tokens and extension to a million.

Vision is native rather than attached. The 27B is a vision-language model with its own encoder, handling images and video from STEM diagrams and scanned documents through to hour-scale footage, and the lab's strongest reported results are on that side of the ledger: 84.3 on OSWorld-Verified for computer use, 81.9 on AndroidWorld for phone use, 64.8 on WebArena-Verified for browser use, 90.0 on MathVision and 91.1 on document parsing. On several of these Alibaba reports it ahead of the much larger previous-generation model it replaces.

Thinking is on by default and can be switched off per request, the effort dial goes from low up to the highest setting, and reasoning from earlier turns is kept in context by default - which Alibaba says both keeps agent decisions consistent and improves cache reuse. The lab adds a caution worth heeding: on multi-turn agentic tasks a lower effort setting does not reliably finish faster, because thinner analysis leads to more failures and retries.

Where it stops is exactly where you would expect the compact member to stop. Alibaba's own text table puts it at 89.2 on GPQA Diamond and 30.8 on HLE, against 92.6 and 43.6 for the Max-class model in the same family, so the ceiling on hard general knowledge is materially lower. Its coding results - 73.0 on Terminal Bench 2.1 and 61.7 on SWE-bench Pro - clear the previous 27B and the previous flagship comfortably but sit below the frontier. The compensation is that it is released under a permissive Apache-2.0 licence, unlike its Max-class sibling.