What Is Qwen3.8-27B? Alibaba's 27B Local Model 2026
The Short Answer
Qwen3.8-27B is Alibaba’s open-weight, dense, multimodal model released August 14, 2026 (15:00 UTC) — eleven days after the trillion-scale flagship Qwen3.8-Max. It’s roughly 27.8B parameters, carries a 262K-token context window, accepts native image and video input, and ships under Apache 2.0 with weights on Hugging Face and ModelScope. Critically, it runs on a single 24GB GPU.
Key Facts
| Qwen3.8-27B | |
|---|---|
| Vendor | Alibaba (Qwen team) |
| Released | August 14, 2026, 15:00 UTC |
| Parameters | ~27.8B, dense (not MoE) |
| Context window | 262,144 tokens |
| Modality | Vision-language: native image + video input |
| License | Apache 2.0 |
| Hardware | Single 24GB consumer GPU (quantized) |
| Weights | Hugging Face (Qwen/Qwen3.8-27B), ModelScope |
| Sibling | Qwen3.8-Max (2.4T params, Aug 3, 2026) |
The Benchmark Story
The generational deltas against Qwen3.6-27B are the reason this release landed hard:
| Benchmark | Qwen3.6-27B | Qwen3.8-27B |
|---|---|---|
| DeepSWE | 13.3 | 42.2 |
| QwenSWEBench | 49.3 | 79.0 |
| Terminal Bench | — | 73.0 |
| OSWorld | — | 84.3 |
A 3× jump on DeepSWE and a 30-point jump on QwenSWEBench in one generation is not a tuning pass — it’s a different training regime for long-horizon agentic work. Alibaba also reports the 27B beating Qwen3.7-Plus, a substantially larger model, on several evals.
Read benchmarks with the usual caution. These are vendor-published numbers, and some harnesses (including SWE-bench Multimodal variants) are judged by other frontier models rather than deterministic tests. Treat them as a signal to evaluate, not as a verdict.
Why the 27B Matters More Than the Flagship
Qwen3.8-Max is a 2.4-trillion-parameter model — impressive, and irrelevant to anyone without a datacenter. The 27B is the release that changes what you can actually run:
- One 24GB card. RTX-class hardware, not H100 clusters.
- Apache 2.0. No revenue-share strings, unlike Qwen3.8-Max’s commercial terms or Kimi K3’s.
- 262K context. Long enough to hold a real codebase slice in a local agent loop.
- Multimodal in. Screenshots and video feed straight into an agent that reads UI state — which is what the 84.3 OSWorld score is measuring.
That combination — permissive license, consumer hardware, agentic scores, vision input — is what local-agent builders have been waiting for.
Where It Fits in the August 2026 Open-Weight Wave
Qwen3.8-27B landed inside a remarkably dense two weeks: Meta’s Muse Glimmer 30B (Aug 10, Apache 2.0), Nvidia’s Nemotron 3.5 Lightning (Aug 11, OpenMDW-1.1), DeepSeek V4 Pro, and Z.ai’s GLM-5.3 (Aug 14). Head-to-head in the local-agent class: Qwen3.8-27B vs Muse Glimmer vs Nemotron 3.5 Lightning.
Who Should Care
- Local-agent builders running always-on loops on a workstation.
- Teams with data-residency constraints who cannot send code to a US or Chinese API.
- Cost engineers routing cheap frequent steps locally and escalating hard steps to a frontier API.
- Anyone building UI-driving agents — the vision input plus OSWorld 84.3 is the differentiator against text-only 30B rivals.
Not for: teams that just want the best result per dollar with no infrastructure work. For that, an API model like Gemini 3.7 Flash or GLM-5.3 is less effort and often cheaper once you price your own GPU time honestly.
Last verified: August 16, 2026.