Model comparison

Qwen3.5-4B vs Nanonets-OCR-s

At Q4_K_M, Nanonets-OCR-s is the smaller download — 1,929,900,800 bytes against 2,707,513,696.

From the file· summed bytes, KV per layer

Side by side

Qwen3.5-4BNanonets-OCR-s
Parameters4.7B3.8B
Architectureqwen35qwen2vl
Layers3236
Native context262,144128,000
Mixture of expertsnono
Quantizations published5756
Smallest quantization1.42 GiB0.74 GiB
Q4_K_M2.52 GiB1.80 GiB
Licenceapache-2.0

KV cache by context

the term that decides long-context viability
ContextQwen3.5-4BNanonets-OCR-sRatio
4,0960.13 GiB0.14 GiB1.12×
8,1920.25 GiB0.28 GiB1.12×
16,3840.50 GiB0.56 GiB1.12×
32,7681.00 GiB1.13 GiB1.12×
65,5362.00 GiB2.25 GiB1.12×
131,0724.00 GiB4.50 GiB1.12×