Qwen3.8-27B
Alibaba · 27B dense
The open multimodal sweet spot at 27B
A 27B dense multimodal model released 14 August 2026 under Apache-2.0: perception 94, documents 94, Chinese 92 and a composite of 90.5. It is not a flagship, but it is among the best current answers to “deployable on one machine, fine-tunable, cleanly licensed”. Video 88 and generation 80 are the boundaries set by its size.
Last updated Oct 21, 2026Editorial scoring
Bottom line
Eight-dimension breakdown
editorial mapping · 0–100- Qwen3.8-27B90.5
Weights: perception 20 / documents 18 / charts 14 / video 14 / agents 14 / generation 6 / Chinese 6 / ecosystem 8. Definitions and sources on the methodology page.
Public benchmarks
as reported, not converted| Benchmark / metric | Value | Note |
|---|---|---|
| 参数规模 | 27B | Dense model |
| 许可证 | Apache-2.0 | Commercial use and fine-tuning |
| 上下文 | 262K | Expandable to 1M |
Public scores come from different versions and prompts and are not interchangeable with our composite; they are one evidence input.
Scenario mapping
five stress tests, one scale| Scenario | Score | Observation |
|---|---|---|
| Chinese receipt OCR | 94 · Stable on print and clean receipts | |
| Chart data extraction | 91 · Reliable on standard charts | |
| 30-min video QA | 86 · Works on sparse frames; misses detail | |
| Mobile GUI control | 86 · Fine for simple tasks | |
| Chinese cultural context | 92 · Solid general context |
Strengths and limitations
Strengths
- 27B dense plus Apache-2.0: the lowest barrier to single-machine deployment and domain fine-tuning.
- Perception 94, documents 94 and Chinese 92 are all above the enterprise floor.
- 262K context, expandable to 1M, reduces chunking for long contracts and many-image inputs.
Limitations
- Generation 80 and video 88: creative work and long video still need a larger model.
- The 27B capacity caps complex agents and hard reasoning.
- Ecosystem 88: tooling is complete, but hosted API pricing and rate limits are less flexible than the Flash tier.
Price and real cost
List price
¥1/M tokens
≈ ¥1 / 1M output tokens; weights free. For 1,000 images at 1,024px, ~1,100 image tokens each and 300 output tokens, the monthly cost is about ¥0.6 (assuming input at 25% of output price).
Use your volumeAbout Qwen3.8-27B
Is Qwen3.8-27B open source?
Licence: Open weights (Apache-2.0). Weights are available, but check the exact terms and filing requirements before commercial use.
How much does Qwen3.8-27B cost?
≈ ¥1 / 1M output tokens; weights free. It ranks #13 with a composite of 90.5. Real bills depend on resolution, tiling, caching and volume discounts.
What is Qwen3.8-27B best for?
Single-machine or small-cluster on-prem; Domain document fine-tuning; Cost-controlled image-text batch work。
Why does Qwen3.8-27B score 90.5?
The composite weights eight dimensions: perception 20%, documents 18%, charts 14%, video 14%, agents 14%, generation 6%, Chinese 6%, ecosystem 8%. Its strongest dimension is Visual perception & grounding; its weakest is Generation & editing.
Personal ranking · Not an official score
Every ranking, score and comparison on this page is an independent personal ranking, not an official score, and represents no vendor, institution or regulator. Scores are a subjective mapping of public evidence under our own weights — change the weights and the order changes. Disagree, or spotted a factual error? Contact the site owner: jizhi2024@foxmail.com — we check each report and log corrections in the changelog.