Is gemma-4-12B-it-qat-q4_0-unquantized or NVIDIA-Nemotron-3-Nano-4B-FP8 smaller?

Both publish quantizations; compare the table for the exact sizes.

Which uses less memory at long context?

gemma-4-12B-it-qat-q4_0-unquantized — its KV cache at 32K context is 2.1x smaller, because of how its attention layers cache.

Model comparison

gemma-4-12B-it-qat-q4_0-unquantized vs NVIDIA-Nemotron-3-Nano-4B-FP8

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: gemma-4-12B-it-qat-q4_0-unquantized's KV cache at 32K is 2.1× 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-unquantized	NVIDIA-Nemotron-3-Nano-4B-FP8
Parameters	12.0B	4.0B
Architecture	gemma4	nemotron_h
Layers	48	42
Native context	262,144	262,144
Mixture of experts	no	no
Quantizations published	2	1
Smallest quantization	6.50 GiB	2.64 GiB
Q4_K_M	—	2.64 GiB
Licence	apache-2.0	other

KV cache by context

the term that decides long-context viability

Context	gemma-4-12B-it-qat-q4_0-unquantized	NVIDIA-Nemotron-3-Nano-4B-FP8	Ratio
4,096	0.72 GiB	0.66 GiB	1.10×
8,192	0.97 GiB	1.31 GiB	1.35×
16,384	1.47 GiB	2.63 GiB	1.79×
32,768	2.47 GiB	5.25 GiB	2.13×
65,536	4.47 GiB	10.50 GiB	2.35×
131,072	8.47 GiB	21.00 GiB	2.48×