AnalysisAI ModelsAugust 26, 2026

Qwen3.8 27B quantizations benchmarked: 4-bit holds up, 1-bit collapses

Q4_K_M quantization of Qwen3.8 27B matches the full BF16 model on Terminal-Bench 2.1 while fitting in 17 GB, but 1-bit performs near random chance on GPQA Diamond. Benchmarks cost about $3,000 on Modal GPUs.

1 source

More stories today

Open the live feed