How much did each quant lose?

Every number here is the drift (KLD) of a quantized model from the original, measured the same way for every publisher and format. Lower is better. Each result carries the hash of the quant file and the pack, so anyone can re-run it.

4 model(s), 33 result(s). Built 2026-11-15T12:04:01Z with quantgap 0.1.1.

Model	Results	Lowest KLD	Pack
Qwen2.5-0.5B-Instruct	15	Q8_0 (0.0025)	`f499c4436fa8e370`
Qwen2.5-1.5B-Instruct	8	Q8_0 (0.0152)	`e2cc88932f748bcb`
Qwen3-0.6B	2	Q3_K_M (0.3953)	`282f68b573fc832a`
SmolLM2-135M-Instruct	8	Q8_0 (0.0047)	`58c854541f94686f`

What is KLD?

At every word position, the original model gives a probability to each possible next token. KLD says how far the quant's probabilities are from those. 0 means identical. We compute it on the reference's top 32 guesses plus one "everything else" bucket, so it is a lower bound of the true KLD. Good for ranking quants, not for comparing with other tools' absolute numbers.