Qwen3-0.6B-3bit-gptq-obq

Qwen3-0.6B quantized to 3-bit (group_size 256, symmetric), 8-bit embedding, ~4.40 bits/weight.

Method: GPTQ within-matrix (act-order) + OBQ across-matrix (GGN + KL-teacher correction). Calibration: gitarist/calibration-generic.

wikitext-2 PPL 36.85, mean KL to fp16 0.694 (fp16 ref PPL 20.96).

Weights are provided dequantized in fp16 for direct loading; quantization is baked in — evaluate as-is.