Qwen3.8-27B Tops 24GB GPU Rankings With Near-Perfect Scores Across Coding, Reasoning, and Document QA
Summary
Qwen3.8-27B dominates 24GB GPU benchmarks with near-perfect scores in coding, reasoning, and document QA, outperforming its predecessor Qwen3.6-27B and rival Gemma 4 31B-it while maintaining identical speed and memory efficiency.
Key Points
- Qwen3.8-27B emerges as the top all-round choice for a 24GB GPU, matching Qwen3.6-27B's ~49 tok/s decode speed and memory footprint while achieving 12/12 coding pass@1, 39/40 reasoning cases, and 23/24 document QA results — a substantial practical upgrade with no meaningful throughput penalty.
- Gemma 4 31B-it proves a strong specialist alternative, posting perfect scores on structured tool calls (90/90 single, 30/30 multi-step) and leading in vision tasks (19/20), but demands higher VRAM, runs ~8.3% slower in text decode, and struggles with iterative coding agent reliability compared to Qwen3.8.
- Qwen3.6-27B remains viable only for maintaining existing validated integrations, as it shows major weaknesses in reasoning (28/120 runs passing), document QA (8/24), and coding reliability versus Qwen3.8 — with identical speed and memory curves offering no practical advantage for fresh deployments.