Model comparison

PaddleOCR-VL-1.6 vs SmolLM2-135M-Instruct

At Q4_K_M, SmolLM2-135M-Instruct is the smaller download — 105,454,144 bytes against 300,219,456. At long context the gap widens: PaddleOCR-VL-1.6's KV cache at 32K is 1.3× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

PaddleOCR-VL-1.6SmolLM2-135M-Instruct
Parameters959M135M
Architecturepaddleocrllama
Layers1830
Native context131,0728,192
Mixture of expertsnono
Quantizations published3641
Smallest quantization0.15 GiB0.08 GiB
Q4_K_M0.28 GiB0.10 GiB
Licenceapache-2.0apache-2.0

KV cache by context

the term that decides long-context viability
ContextPaddleOCR-VL-1.6SmolLM2-135M-InstructRatio
4,0960.07 GiB0.09 GiB1.25×
8,1920.14 GiB0.18 GiB1.25×
16,3840.28 GiB0.35 GiB1.25×
32,7680.56 GiB0.70 GiB1.25×
65,5361.13 GiB1.41 GiB1.25×
131,0722.25 GiB2.81 GiB1.25×