Is MathCoder2-CodeLlama-7B or gemma-4-E4B-it smaller?

MathCoder2-CodeLlama-7B is smaller at Q4_K_M: 3.80 GiB against 4.64 GiB.

Which uses less memory at long context?

gemma-4-E4B-it — its KV cache at 32K context is 31.6x smaller, because of how its attention layers cache.

Model comparison

MathCoder2-CodeLlama-7B vs gemma-4-E4B-it

At Q4_K_M, MathCoder2-CodeLlama-7B is the smaller download — 4,081,096,640 bytes against 4,977,171,584. At long context the gap widens: gemma-4-E4B-it's KV cache at 32K is 31.6× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	MathCoder2-CodeLlama-7B	gemma-4-E4B-it
Parameters	6.7B	8.0B
Architecture	llama	gemma4
Layers	32	42
Native context	16,384	131,072
Mixture of experts	no	no
Quantizations published	20	36
Smallest quantization	2.20 GiB	3.30 GiB
Q4_K_M	3.80 GiB	4.64 GiB
Licence	apache-2.0	apache-2.0

KV cache by context

the term that decides long-context viability

Context	MathCoder2-CodeLlama-7B	gemma-4-E4B-it	Ratio
4,096	2.00 GiB	0.12 GiB	16.25×
8,192	4.00 GiB	0.18 GiB	22.51×
16,384	8.00 GiB	0.29 GiB	27.86×
32,768	16.00 GiB	0.51 GiB	31.63×
65,536	32.00 GiB	0.94 GiB	33.92×
131,072	64.00 GiB	1.82 GiB	35.20×