Model comparison

Apriel-1.6-15b-Thinker vs gemma-4-31B-it

At Q4_K_M, Apriel-1.6-15b-Thinker is the smaller download — 8,785,478,144 bytes against 18,323,733,440.

From the file· summed bytes, KV per layer

Side by side

Apriel-1.6-15b-Thinkergemma-4-31B-it
Parameters14.9B31.3B
Architecturellamagemma4
Layers4860
Native context262,400262,144
Mixture of expertsnono
Quantizations published6055
Smallest quantization3.22 GiB7.95 GiB
Q4_K_M8.18 GiB17.07 GiB
Licencemitapache-2.0

KV cache by context

the term that decides long-context viability
ContextApriel-1.6-15b-Thinkergemma-4-31B-itRatio
4,0960.75 GiB1.80 GiB2.40×
8,1921.50 GiB2.42 GiB1.61×
16,3843.00 GiB3.67 GiB1.22×
32,7686.00 GiB6.17 GiB1.03×
65,53612.00 GiB11.17 GiB1.07×
131,07224.00 GiB21.17 GiB1.13×