Model comparison

LightOnOCR-2-1B-bbox-soup vs Llama-3.2-3B-Instruct

At Q4_K_M, LightOnOCR-2-1B-bbox-soup is the smaller download — 396,701,728 bytes against 2,019,377,440.

From the file· summed bytes, KV per layer

Side by side

LightOnOCR-2-1B-bbox-soupLlama-3.2-3B-Instruct
Parameters1.0B3.2B
Architectureqwen3llama
Layers2828
Native context16,384131,072
Mixture of expertsnono
Quantizations published839
Smallest quantization0.34 GiB0.85 GiB
Q4_K_M0.37 GiB1.88 GiB
Licence

KV cache by context

the term that decides long-context viability
ContextLightOnOCR-2-1B-bbox-soupLlama-3.2-3B-InstructRatio
4,0960.44 GiB0.44 GiB1.00×
8,1920.88 GiB0.88 GiB1.00×
16,3841.75 GiB1.75 GiB1.00×
32,7683.50 GiB3.50 GiB1.00×
65,5367.00 GiB7.00 GiB1.00×
131,07214.00 GiB14.00 GiB1.00×