Firsthand benchmarks of Qwen 3.8-27B against 3.6-27B on one RTX 3090: generation within a percent, VRAM +254 MiB, and HumanEval pass@1 a statistical tie.