Model comparison

gemma-4-12B-it-qat-q4_0-unquantized vs Mellum2-12B-A2.5B-Instruct

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: Mellum2-12B-A2.5B-Instruct's KV cache at 32K is 4.9× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

gemma-4-12B-it-qat-q4_0-unquantizedMellum2-12B-A2.5B-Instruct
Parameters12.0B12.1B
Architecturegemma4mellum
Layers4828
Native context262,144131,072
Mixture of expertsnoyes, 64 experts
Quantizations published22
Smallest quantization6.50 GiB7.52 GiB
Q4_K_M7.52 GiB
Licenceapache-2.0apache-2.0

KV cache by context

the term that decides long-context viability
Contextgemma-4-12B-it-qat-q4_0-unquantizedMellum2-12B-A2.5B-InstructRatio
4,0960.72 GiB0.12 GiB6.18×
8,1920.97 GiB0.17 GiB5.67×
16,3841.47 GiB0.28 GiB5.24×
32,7682.47 GiB0.50 GiB4.95×
65,5364.47 GiB0.94 GiB4.77×
131,0728.47 GiB1.81 GiB4.67×