Model comparison

Supra-50M-Base vs tinygemma3_cifar

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: tinygemma3_cifar's KV cache at 32K is 3.0× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

Supra-50M-Basetinygemma3_cifar
Parameters52M39M
Architecturellamagemma3
Layers128
Native context1,024131,072
Mixture of expertsnono
Quantizations published331
Smallest quantization0.02 GiB0.04 GiB
Q4_K_M0.03 GiB
Licenceapache-2.0wtfpl

KV cache by context

the term that decides long-context viability
ContextSupra-50M-Basetinygemma3_cifarRatio
4,0960.05 GiB0.06 GiB1.33×
8,1920.09 GiB0.08 GiB1.22×
16,3840.19 GiB0.09 GiB2.02×
32,7680.38 GiB0.12 GiB3.02×
65,5360.75 GiB0.19 GiB4.02×
131,0721.50 GiB0.31 GiB4.82×