Is LFM2-2.6B or Llama-3.1-8B-Instruct smaller?

LFM2-2.6B is smaller at Q4_K_M: 1.46 GiB against 4.58 GiB.

Which uses less memory at long context?

LFM2-2.6B — its KV cache at 32K context is 8.0x smaller, because of how its attention layers cache.

Model comparison

LFM2-2.6B vs Llama-3.1-8B-Instruct

At Q4_K_M, LFM2-2.6B is the smaller download — 1,563,668,704 bytes against 4,920,739,168. At long context the gap widens: LFM2-2.6B's KV cache at 32K is 8.0× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	LFM2-2.6B	Llama-3.1-8B-Instruct
Parameters	2.6B	8.0B
Architecture	lfm2	llama
Layers	30	32
Native context	128,000	131,072
Mixture of experts	no	no
Quantizations published	30	45
Smallest quantization	1.05 GiB	2.02 GiB
Q4_K_M	1.46 GiB	4.58 GiB
Licence	other	llama3.1

KV cache by context

the term that decides long-context viability

Context	LFM2-2.6B	Llama-3.1-8B-Instruct	Ratio
4,096	0.06 GiB	0.50 GiB	7.99×
8,192	0.13 GiB	1.00 GiB	8.00×
16,384	0.25 GiB	2.00 GiB	8.00×
32,768	0.50 GiB	4.00 GiB	8.00×
65,536	1.00 GiB	8.00 GiB	8.00×
131,072	2.00 GiB	16.00 GiB	8.00×