Qwen3.8-27B
27B native multimodal dense — vendor: beats 3.7-Plus overall
Official @Alibaba_Qwen on 2026-08-14: a 27B native multimodal dense model. Vendor says it outperforms Qwen3.7-Plus overall. Native context 262K, YaRN to 1M. Apache 2.0. Hugging Face repo Qwen/Qwen3.8-27B. This is a local / open-weight page, not a hosted SKU page. QCode does not list this id.
Local-first facts
Dense multimodal
Native multimodal dense model, not the 2.4T MoE. Official @Alibaba_Qwen.
Native context
Native 262K. 1M needs YaRN. It is not a native 1M window.
YaRN stretch
Official post: YaRN can take it to 1M. Do not write that as native.
License
Open weights. HF path Qwen/Qwen3.8-27B.
What the 27B is
Qwen3.8-27B is the open-weight drop from official @Alibaba_Qwen on 2026-08-14: a 27B native multimodal dense model for people who will run it themselves. It is not hosted Max and not 2.4T-A95B. Repo: Hugging Face Qwen/Qwen3.8-27B.
“Beats 3.7-Plus” is the vendor line
@Alibaba_Qwen says it outperforms Qwen3.7-Plus overall. That is a vendor claim; this page does not invent an independent table. Native 262K, YaRN to 1M, and Apache 2.0 can be checked against the official post and the repo.
Timeline
3.8-Max ships and open weights are promised.
@Alibaba_Qwen drops Qwen3.8-27B and Qwen3.8-2.4T-A95B.
Weights are downloadable. QCode /models has neither the 27B nor qwen3.8-max. For API use qwen3.7-max.
Confirmed vs treat carefully
Confirmed
Official post 2026-08-14; 27B native multimodal dense; native 262K, YaRN to 1M; Apache 2.0; HF Qwen/Qwen3.8-27B. Vendor: outperforms Qwen3.7-Plus overall.
Treat carefully
“Beats 3.7-Plus” is a vendor claim, not a rerun on this page. Do not write YaRN 1M as native 1M. Do not treat 27B as a QCode id — it is not listed.
Local vs API
Run the 27B yourself
Want weights, a clean license, and your own runtime? Pull Qwen/Qwen3.8-27B. Apache 2.0.
Call Qwen on QCode
QCode does not list the 27B. For Qwen on the platform, use listed qwen3.7-max.
How to run it locally
Pull weights from Hugging Face Qwen/Qwen3.8-27B (or the ModelScope twin). Size the runtime for native 262K. Turn on YaRN only if you need 1M, following the official recipe — do not assume every engine enables it. This is a self-host path.
On QCode
QCode /models does not list Qwen3.8-27B, and it does not list qwen3.8-max. To call Qwen via API on QCode today, use qwen3.7-max. The 27B is open weights, not a hosted id here.
Qwen3.8-27B FAQ
What is the 27B?
A 27B native multimodal dense model. Official @Alibaba_Qwen, 2026-08-14.
Does it beat 3.7-Plus?
The vendor says it outperforms Qwen3.7-Plus overall. We do not invent independent scores.
How long is context?
Native 262K. YaRN can go to 1M — that is not the native window.
License?
Apache 2.0. HF: Qwen/Qwen3.8-27B.
Can I call the 27B on QCode?
No. That id is not in the catalog.
Which API id then?
For Qwen on QCode, use qwen3.7-max.
Sources
Official @Alibaba_Qwen 2026-08-14; Hugging Face Qwen/Qwen3.8-27B. “Beats 3.7-Plus overall” is a vendor claim.
Run 27B locally. API goes to 3.7-Max
The 27B is not listed. qwen3.7-max is.
Related
Qwen3.8-Max guide
The 2.4T hosted flagship. Also not on QCode.
Qwen3.7-Max guide
The Qwen API QCode lists: qwen3.7-max.
2026 model radar
Which ids are actually in the catalog.
Not affiliated with Alibaba / Qwen. QCode does not offer Qwen3.8-27B. “Outperforms 3.7-Plus” is a vendor claim.