27B open weights ship under Apache 2.0

Alibaba has published the weights of Qwen3.8-27B on Hugging Face under the Apache 2.0 license, a 27-billion-parameter dense model with a 262K native context window. The release was announced in a post on Threads, which describes the model as natively multimodal and runnable locally.

At a glance

• The context window starts at 262K tokens natively and can be pushed up to 1 million via YaRN, the context extension method named in the release post.
• According to the release post, the 27B model performs on par with Opus 4.6 on benchmarks, a claim published by Alibaba itself and one the open weights now make verifiable outside the company.
• Alibaba published the 27B weights alongside Qwen3.8-2.4T-A95B, the Max-class version of the same generation, so the family now spans a locally runnable dense model and a much larger sibling.

Why it matters A 27B dense model with permissive terms sits in the size class that matters most for teams running inference on their own hardware, and Apache 2.0 removes the licensing review that slows adoption of restricted open weights. The parity claim against Opus 4.6 reads as positioning rather than a settled result, but it is now falsifiable: anyone can run the same evaluations locally and publish what they find.

Qwen3.8-27B is a native multimodal dense model. Dense means every parameter is active on each forward pass, unlike mixture-of-experts designs that route each token through a subset of the network. The release post presents it as a model that runs on local hardware.

The native context window is 262K tokens, and the release post says it can be extended to 1 million with YaRN, a method for stretching a model's trained positional range beyond its training length. That upper figure is an extension result rather than the window the model was trained on.

The weights ship under Apache 2.0, among the most permissive terms used for open model releases. The license allows commercial use, modification and redistribution of derivatives, and attaches no user thresholds or field-of-use restrictions of the kind carried by several other open-weight licenses.

Alibaba had signalled the open release a week earlier and published the 27B weights alongside Qwen3.8-2.4T-A95B, the Max-class version of the same generation. The 27B model is the one presented as locally runnable, while the Max-class release sits above it in the same family.

What's next The open license makes the Opus 4.6 parity claim checkable outside Alibaba, since the weights are downloadable and independent evaluations of a dense model in this size class usually follow a Hugging Face release quickly. What remains unclear is how the 1M-token YaRN extension behaves in practice and what memory it demands on local hardware. Deployment terms for Qwen3.8-2.4T-A95B are not specified.