Is CodeLlama-7b-instruct-hf or gemma-4-12B-it-qat-q4_0-unquantized 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 6.5x smaller, because of how its attention layers cache.

Model comparison

CodeLlama-7b-instruct-hf vs gemma-4-12B-it-qat-q4_0-unquantized

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 6.5× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	CodeLlama-7b-instruct-hf	gemma-4-12B-it-qat-q4_0-unquantized
Parameters	6.7B	12.0B
Architecture	llama	gemma4
Layers	32	48
Native context	16,384	262,144
Mixture of experts	no	no
Quantizations published	12	2
Smallest quantization	2.63 GiB	6.50 GiB
Q4_K_M	3.80 GiB	—
Licence	llama2	apache-2.0

KV cache by context

the term that decides long-context viability

Context	CodeLlama-7b-instruct-hf	gemma-4-12B-it-qat-q4_0-unquantized	Ratio
4,096	2.00 GiB	0.72 GiB	2.78×
8,192	4.00 GiB	0.97 GiB	4.13×
16,384	8.00 GiB	1.47 GiB	5.45×
32,768	16.00 GiB	2.47 GiB	6.48×
65,536	32.00 GiB	4.47 GiB	7.16×
131,072	64.00 GiB	8.47 GiB	7.56×