MODEL PROFILE / EDITION 2026.09
Qwen3.8-27B
Teams that need a self-hostable, screenshot-capable frontend model on their own hardware
Balanced model evidence score; rank 17. No rank delta is shown because there is no prior numeric snapshot.
Best for
Teams that need a self-hostable, screenshot-capable frontend model on their own hardware
Price & access
Hosted $0.50 in / $3.00 out per 1M (Singapore/international); $0.424/$1.696 in China (Beijing); a third-party comparison lists $0.425/$2.55; self-hosting has no token price
Evidence summary
Detailed vendor model card with harness disclosure, Alibaba pricing, Arena and AA. Licence is only confirmed by secondary sources, and release date differs between QwenCloud (Aug 19) and AA (Aug 14).
Edition status
No prior snapshot. Tracked in the September 2026 baseline.
Strengths
- Frontier-adjacent Arena frontend Elo (1601) from a 27-28B dense model
- Apache-2.0 licensing with no field-of-use restriction (per AA and third-party comparison)
- Vision2Web 62.9 plus OSWorld-Verified 84.3 and AndroidWorld 81.9 - unusually strong UI grounding for its size
- Realistically self-hostable on a single well-specified node
Tradeoffs
- DeepSWE 1.1 42.2 shows a real gap to frontier models on long-horizon SWE work
- 43.9 tok/s output speed is slow
- The Hugging Face card does not state a licence - Apache 2.0 comes from secondary sources
- Effort setting swings its AA index from 26 to 41
Score profile
visual 73 · code 76 · agentic 76 · debug 70 · context 80 · speed 66 · value 90
Linked evidence
Evidence source 1 ↗Evidence source 2 ↗Evidence source 3 ↗Evidence source 4 ↗Evidence source 5 ↗Evidence source 6 ↗Evidence source 7 ↗Use the same published data, filters, and side-by-side metrics.