$ cat wiki/models/qwen-3-8-27b.md

Qwen 3.8 27B

modelupdated 2026-08-04created 2026-08-04

chinese-lab open-weight-announced local-inference announced-not-released

Spec

Attribute	Value
Developer	Alibaba / Qwen AI Lab
Released	not yet
Announced	2026-08-03
Context window	unknown
Pricing	unknown
License	Open-weight announced for 2026-08-10; not yet shipped
Availability	not yet available — Alibaba Cloud Model Studio stated
Almost every row here is `unknown`, and that is the accurate state of this model
rather than a gap in the reading. The announcement named the checkpoint, its intended
audience and an open-weights date; it did not publish a parameter count, an
architecture (dense or MoE), a context window, a price or a licence name
(source).

The 27B in the name is Alibaba's, not a disclosed parameter count confirmed by this wiki — the sibling Qwen 3.8 Max carries 2.4T total / 95B active, so the naming convention on this line is not a reliable guide to what the number refers to.

Release Date

Announced 2026-08-03 alongside the general availability of Qwen 3.8 Max. Open weights stated for 2026-08-10 (source).

Not released as of this page's creation on 2026-08-04.

Benchmarks

None published. No first-party or third-party measurement of this checkpoint was found in any source read (source). It appears in neither the Artificial Analysis table read 2026-08-02 (source) nor the LMArena snapshot of the same date (source), both of which predate the announcement.

Use Cases

Alibaba's stated positioning is on-premise deployment: this is described as the checkpoint "that fits ordinary on-premise GPU hardware", which organisations "can run on infrastructure they own and fine-tune on their own data" (source).

Third-party hardware claim, recorded as a claim. Unsloth states the model "will run locally on 17GB RAM/VRAM setups" (source). This is a statement about an unreleased checkpoint made before any weights existed to measure, and it is the only quantitative figure attached to this model anywhere in this wiki. It is not independent confirmation of anything.

Compared To

Qwen 3.8 Max — the flagship announced the same day; 2.4T total / 95B active, API-first, open weights on the same stated date
Inkling — Thinking Machines' 975B/41B Apache 2.0 release, the nearest Western comparison by intent
Laguna S 2.1 — Poolside's 118B/8B OpenMDW-1.1 release
Kimi K3 — Moonshot's 2.8T open-weight model, the Chinese release this one is implicitly positioned against on deployability rather than scale