Qwen3-0.6B-3bit-qqq-gptq-obq

Qwen3-0.6B, weights quantized to 3-bit (group_size 128, symmetric), 8-bit embedding, ~4.48 bits/weight.

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

wikitext-2 PPL 30.54, mean KL to fp16 0.516 (fp16 ref PPL 20.96).

With per-token dynamic int8 activations (W3A8): PPL 30.82, KL 0.525.

Weights are provided dequantized in fp16; weight quantization is baked in. The W3A8 numbers require per-token int8 activation quantization applied at inference.