Is SmolLM3-3B or Llama-3.1-8B-Instruct smaller?

SmolLM3-3B is smaller at Q4_K_M: 1.78 GiB against 4.58 GiB.

Which uses less memory at long context?

SmolLM3-3B — its KV cache at 32K context is 1.8x smaller, because of how its attention layers cache.

Model comparison

SmolLM3-3B vs Llama-3.1-8B-Instruct

At Q4_K_M, SmolLM3-3B is the smaller download — 1,915,305,312 bytes against 4,920,739,168. At long context the gap widens: SmolLM3-3B's KV cache at 32K is 1.8× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	SmolLM3-3B	Llama-3.1-8B-Instruct
Parameters	3.1B	8.0B
Architecture	smollm3	llama
Layers	36	32
Native context	65,536	131,072
Mixture of experts	no	no
Quantizations published	44	45
Smallest quantization	0.89 GiB	2.02 GiB
Q4_K_M	1.78 GiB	4.58 GiB
Licence	apache-2.0	llama3.1

KV cache by context

the term that decides long-context viability

Context	SmolLM3-3B	Llama-3.1-8B-Instruct	Ratio
4,096	0.28 GiB	0.50 GiB	1.78×
8,192	0.56 GiB	1.00 GiB	1.78×
16,384	1.13 GiB	2.00 GiB	1.78×
32,768	2.25 GiB	4.00 GiB	1.78×
65,536	4.50 GiB	8.00 GiB	1.78×
131,072	9.00 GiB	16.00 GiB	1.78×