Is Amaretto-3B or Llama-3.2-1B-Instruct smaller?

Llama-3.2-1B-Instruct is smaller at Q4_K_M: 0.75 GiB against 2.00 GiB.

Which uses less memory at long context?

Llama-3.2-1B-Instruct — its KV cache at 32K context is 3.3x smaller, because of how its attention layers cache.

Model comparison

Amaretto-3B vs Llama-3.2-1B-Instruct

At Q4_K_M, Llama-3.2-1B-Instruct is the smaller download — 807,690,688 bytes against 2,146,498,240. At long context the gap widens: Llama-3.2-1B-Instruct's KV cache at 32K is 3.3× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	Amaretto-3B	Llama-3.2-1B-Instruct
Parameters	4.3B	1.2B
Architecture	mistral3	llama
Layers	26	16
Native context	262,144	131,072
Mixture of experts	no	no
Quantizations published	36	39
Smallest quantization	0.87 GiB	0.39 GiB
Q4_K_M	2.00 GiB	0.75 GiB
Licence	apache-2.0	—

KV cache by context

the term that decides long-context viability

Context	Amaretto-3B	Llama-3.2-1B-Instruct	Ratio
4,096	0.41 GiB	0.13 GiB	3.25×
8,192	0.81 GiB	0.25 GiB	3.25×
16,384	1.63 GiB	0.50 GiB	3.25×
32,768	3.25 GiB	1.00 GiB	3.25×
65,536	6.50 GiB	2.00 GiB	3.25×
131,072	13.00 GiB	4.00 GiB	3.25×