Unsloth's new Qwen3.8 quants gain 5–10% accuracy at the same size

Unsloth Qwen3.8-27B Quantizations – Dynamic v3.1 Update

Unsloth released updated Qwen3.8-27B GGUFs using its Dynamic V3.1 algorithm, which boosts average accuracy by 5–10% at comparable sizes across Terminal Bench 2.1, DeepSWE 1.1, LiveCodeBench v6, GPQA Diamond, and other benchmarks. The 27B dense model offers vision, reasoning, a 256K context window, and runs on 17GB RAM/VRAM. Unsloth also disputes ISTA's GSQ-RCO benchmark claims, showing its older UD3 quants outperform ISTA's on average over 14 trials versus one.