Alibaba Put a 27B Model Under Apache 2.0 and Says It Beats Its Own Bigger One

Alibaba's Qwen team published the weights for Qwen3.8-27B on 14 August, releasing them on Hugging Face and ModelScope under the Apache 2.0 licence.

What the model is

A 27B-parameter dense vision-language model — not a mixture-of-experts — covering text, images and multi-hour video. The architecture is a hybrid across 64 layers: sixteen blocks of three Gated DeltaNet layers feeding a network, each followed by a single Gated Attention block. Native context is 262,144 tokens, extensible to roughly 1M with YaRN, and the thinking mode can be switched off.

The claim

Qwen says the 27B beats the substantially larger Qwen3.7-Plus on coding and office work. Its card reports SWE-bench Pro 61.7, Terminal Bench 2.1 73.0, OSWorld-Verified 84.3, GPQA Diamond 89.2, MathVision 94.6 with code interpreter, and CharXiv 90.2.

Read the licence precisely

Apache 2.0 here is real: the weights are downloadable and commercially usable without a bespoke licence. But training data and training code are not released, so this is open weights, not open source — a distinction that matters for anyone who needs to audit what the model learned rather than just run it.

And read the benchmarks the same way

Every figure above is self-reported by Alibaba on its own model card, with no independent replication yet. A small dense model beating a much larger sibling is exactly the comparison that tends to shrink under third-party evaluation, because the vendor chooses both the harness and the baseline. The official API is listed as coming soon, and a hosted 1M-context version is promised on Qwen Cloud.

What is actually new here

The Qwen 3.8 name is not. It was used in July for the closed Qwen3.8-Max, and the much larger open-weight Qwen3.8-2.4T-A95B mixture-of-experts model was posted two days earlier. The news is that the small, dense, locally runnable tier of the family is now out — which is the tier most people can actually deploy.