Qwen: Qwen3.8 27B

The superlative on this model's card is not about this model. The Hugging Face page for Qwen/Qwen3.8-27B introduces Qwen3.8, the most capable generation in the Qwen open-model family to datesource, accessed 2026-08-28 — the subject is the Qwen3.8 generation, not the 27B row carrying the sentence. Within that generation the open lead belongs elsewhere: qwen/qwen3.8-2.4t-a95b, listed two days before this row, is open by the same test — it carries a Hugging Face id where the cloud-only qwen/qwen3.8-max carries none — and in the 31 August 2026 snapshot it outscores this row on the Artificial Analysis intelligence index. A reader who takes the card's line as a claim about the 27B is reading past the model that beats it.

What is genuinely new here is the window. One generation earlier and at the same size, qwen/qwen3.6-27b listed 262144tokensopenrouter-models, last checked 2026-09-14 of context; this row lists 1000000tokensopenrouter-models, last checked 2026-09-14. Every Qwen release with published weights before August 2026 stops at the smaller of those two figures — the ceiling held from July 2025 to August 2026, across more than thirty rows with published weights in the catalog as observed on 31 August 2026, and then broke twice, two days apart: once on qwen/qwen3.8-2.4t-a95b at 1048576tokensopenrouter-models, last checked 2026-09-14, and once here, on a model of 27Bsource, accessed 2026-08-28.

The licence is what makes that size matter. Alibaba Cloud's record of its policy reads Apache License 2.0 for open releases; a source-available Qwen License on larger modelssource, accessed 2026-08-28, and this row sits on the permissive side of the split: Apache License 2.0source, accessed 2026-08-28, which carries no field-of-use and no user-count clause. Against the closed flagship's not publishedlast known value, as of an unrecorded date — the source no longer lists this rowopenrouter-models, last checked 2026-09-14 input this row lists $0.21per million tokensopenrouter-models, last checked 2026-09-14 — and on self-hosted weights that per-token rate is optional entirely.