Is Nanonets-OCR-s or LFM2.5-VL-1.6B smaller?

LFM2.5-VL-1.6B is smaller at Q4_K_M: 0.68 GiB against 1.80 GiB.

Which uses less memory at long context?

LFM2.5-VL-1.6B — its KV cache at 32K context is 3.0x smaller, because of how its attention layers cache.

Model comparison

Nanonets-OCR-s vs LFM2.5-VL-1.6B

At Q4_K_M, LFM2.5-VL-1.6B is the smaller download — 730,895,968 bytes against 1,929,900,800. At long context the gap widens: LFM2.5-VL-1.6B's KV cache at 32K is 3.0× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	Nanonets-OCR-s	LFM2.5-VL-1.6B
Parameters	3.8B	1.6B
Architecture	qwen2vl	lfm2
Layers	36	16
Native context	128,000	128,000
Mixture of experts	no	no
Quantizations published	56	23
Smallest quantization	0.74 GiB	0.45 GiB
Q4_K_M	1.80 GiB	0.68 GiB
Licence	—	other

KV cache by context

the term that decides long-context viability

Context	Nanonets-OCR-s	LFM2.5-VL-1.6B	Ratio
4,096	0.14 GiB	0.05 GiB	3.00×
8,192	0.28 GiB	0.09 GiB	3.00×
16,384	0.56 GiB	0.19 GiB	3.00×
32,768	1.13 GiB	0.38 GiB	3.00×
65,536	2.25 GiB	0.75 GiB	3.00×
131,072	4.50 GiB	1.50 GiB	3.00×