Qwen3.8-27B
Qwen's 27B vision-language dense ships Apache-2.0 — with the LICENSE file its coder cousin still owes us
What it is
The open-weights anchor of the Qwen3.8 generation: a 27B dense vision-language model mixing Gated DeltaNet and Gated Attention across 64 layers, with 262K native context extensible to 1M and image and video input alongside text. It's the consumer-hardware release of the pair announced with the much larger Qwen3.8-Max, shipped on the official Qwen org with FP8 builds and day-one GGUF support across the ecosystem.
Why it's interesting
The pre-release license anxiety — Apache or Tongyi Qianwen? — resolved the right way: an Apache-2.0 tag AND a standalone 11.5 kB LICENSE file in the repo root. From the org that left Qwen3-Coder-Next in our missing-LICENSE club, a full Apache text alongside the weights is worth naming. The uptake is immediate: ~267k downloads and five-digit likes within days.
Use cases
- Local multimodal assistants that read screenshots, documents, and video
- Long-context RAG and document pipelines within 262K native tokens
- Fine-tuning a current-gen dense VLM without license negotiation
Who it's for
Local-AI users with a serious GPU, VLM fine-tuners, multimodal app builders
Setup
Moderate. ~56GB BF16 across 18 shards; FP8 official and community GGUF quants bring it to single-GPU and high-memory Mac territory
Limitations & cautions
27B dense wants real hardware before quantization, the 1M extended context relies on scaling tricks beyond the 262K native window, and launch benchmarks are Alibaba's own — the larger Qwen3.8-Max sibling's license was still undisclosed at check time.
Editorial takeaway
Same org, two repos: the coder ships a tag and a promise, this one ships the license. The file is the policy — everything else is a press release.