Qwen3.8-27B
modelYour notes
The dense member of the Qwen3.8 generation, released August 14, 2026 under Apache 2.0. Unlike the 2.4T-A95B flagship, which ships under a custom license, it carries a standard open-source license. It is a native vision-language model with image and video input. The language model keeps the Qwen3.5 hybrid layout: 64 layers arranged as 16 × (3 Gated DeltaNet blocks, then 1 gated attention block), a 5120 hidden size, a 248,320-token vocabulary and multi-step MTP. Context is 262,144 tokens natively and extends to 1M. Thinking is on by default and can be switched off per request, with reasoning_effort control and preserved thinking across turns. Its HuggingFace repo counted about 7.5M downloads over the last month in late October.
Qwen's card compares it with Qwen3.6-27B, Qwen3.7-Plus and Claude Opus 4.6 (max). It reports SWE-bench Pro 61.7 against Opus 4.6's 53.4 (Qwen's Claude Code harness for every model except Opus, whose figure is Anthropic's own), Terminal-Bench 2.1 73.0 (Opus 78.2), GPQA Diamond 89.2, LiveCodeBench v6 90.3 and IFBench 79.5. On vision agents it reports OSWorld-Verified 84.3 and AndroidWorld 81.9. All are self-reported.
Artificial Analysis scores it 20 on Intelligence Index v5.0 at xhigh effort, down from 34 on v4.3.2 and 52 on v4.1.1. Its lower effort settings had not yet been rescored on v5 at launch, and read 28 (medium), 26 (low) and 22 (non-reasoning) on v4.3.2. The v5 drop is steepest on the agentic and science evals: 5.6% on Terminal-Bench 4.0, 0.5% on Terminal-Bench-Science and 4% on the CritPt v2 hard subset, against 48.2% on AutomationBench-AA and 14.9% on FrontierCode 2.0. Its cost per index task is $3.53, at $0.50 / $3 per million tokens.
Model Details
Benchmark Scores
| Benchmark | Score | Mode |
|---|---|---|
| SWE-bench Pro | 61.7 | — |
| Terminal-Bench 2.1 | 73.0 | — |
| GPQA Diamond | 89.2 | — |
| LiveCodeBench v6 | 90.3 | — |
| OSWorld-Verified | 84.3 | — |
Variants
| Name | Parameters | Notes |
|---|---|---|
| Qwen3.8-27B (xhigh) | — | AA Intelligence Index v5.0 20 (top-level anchor; 34 on v4.3.2) |
| Qwen3.8-27B (medium / low / non-reasoning) | — | AA Intelligence Index v4.3.2 28 / 26 / 22; not yet rescored on v5.0 |