SmolLM2-135M-Instruct

pack SmolLM2-135M-Instruct@58c854541f94686f (id 58c854541f94686f), 8 result(s). Lower KLD is better. Click a column to sort.

KLD mean against file size. Bars are 95% intervals. Lower and further left is better.

Quant	Publisher	Size	KLD mean	95% interval	Worst 1%	Top-1	PPL ratio	Loops (q/ref)	JSON ok	Backend
Q8_0	bartowski	144.8 MB	0.0047	0.0045 to 0.0050	0.030	95.1%	1.003	27/26	3/8	llamacpp
Q6_K	bartowski	138.4 MB	0.0066	0.0062 to 0.0069	0.046	94.3%	1.005	25/26	2/8	llamacpp
Q5_K_M	bartowski	112.1 MB	0.0168	0.0156 to 0.0179	0.130	91.5%	1.014	26/26	6/8	llamacpp
Q4_K_M	bartowski	105.5 MB	0.0361	0.0340 to 0.0382	0.267	87.8%	1.036	30/26	2/8	llamacpp
Q3_K_M	bartowski	93.5 MB	0.0760	0.0711 to 0.0808	0.598	83.1%	1.073	36/26	0/8	llamacpp
IQ4_XS	bartowski	90.9 MB	0.0911	0.0854 to 0.0971	0.747	81.9%	1.087	37/26	5/8	llamacpp
IQ3_M	bartowski	90.2 MB	0.1263	0.1185 to 0.1351	1.028	78.8%	1.121	42/26	3/8	llamacpp
Q2_K	bartowski	88.2 MB	0.1871	0.1757 to 0.1998	1.374	75.3%	1.191	36/26	4/8	llamacpp

approx = the server only reported its own top-N, so some reference tokens were counted in the tail bucket. nd = the same text gave slightly different numbers on two runs. "Loops" shows quant / original. All results used the same probe set and pack.