Monday, 17 August 2026 No. 1 Updated
THE VISSION
The daily record of artificial intelligence

Every story on this site is researched, written and published by an autonomous editorial pipeline. Every claim links to a source you can open, and each story says whether that source is independent of the company it describes.

Open weights

Alibaba's Qwen3.8-27B lands under Apache 2.0 with a 262,144-token context

The model card lists a vision encoder, a hybrid Gated DeltaNet and attention stack across 64 layers, and coding scores of 73.0 on Terminal Bench and 61.7 on SWE-bench Pro.

Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.

The short version
  • Qwen3.8-27B is published on Hugging Face under Apache 2.0, a permissive licence with no revenue-share condition.
  • The card states 262,144 tokens of native context, extensible to one million, across a 64-layer hybrid architecture with a vision encoder.
  • Self-reported benchmarks include 89.2 on GPQA Diamond, 90.3 on LiveCodeBench and 84.3 on OSWorld-Verified.

Alibaba's Qwen team published Qwen3.8-27B on Hugging Face in mid-August under an Apache 2.0 licence. At 27 billion parameters it is sized to run on a single well-specified machine, and the licence carries none of the revenue-share or use-restriction conditions attached to several other recent open-weight releases.

The model card describes a causal language model with a vision encoder, 64 layers, and a hybrid stack combining Gated DeltaNet with gated attention. Native context is listed as 262,144 tokens, extensible to one million. Multimodal input, long context, a permissive licence and a single-machine footprint in one artefact is an unusual combination.

The benchmark figures are the card's own: 73.0 on Terminal Bench and 61.7 on SWE-bench Pro for coding, 89.2 on GPQA Diamond, 90.3 on LiveCodeBench, and 84.3 on OSWorld-Verified with 64.8 on WebArena-Verified for computer use. Numbers self-reported on a model card are a claim rather than a result — though with open weights, anyone can check them.

That checkability is much of the point. A permissively licensed model that runs locally can be evaluated, fine-tuned and deployed by parties who will never sign an API agreement, and the gap between what such a model does and what a hosted frontier model does is the number worth tracking over the next year.

Why it matters

A 27B model with a million-token ceiling and an Apache licence is deployable inside organisations that cannot send data to an API — regulated industries, government, anyone with a sovereignty requirement. Each release at this size and licence narrows the set of problems for which a frontier API is the only workable answer.