Is PaddleOCR-VL-1.6 or Qwen2.5-0.5B-Instruct smaller?

PaddleOCR-VL-1.6 is smaller at Q4_K_M: 0.28 GiB against 0.37 GiB.

Which uses less memory at long context?

Qwen2.5-0.5B-Instruct — its KV cache at 32K context is 1.5x smaller, because of how its attention layers cache.

Model comparison

PaddleOCR-VL-1.6 vs Qwen2.5-0.5B-Instruct

At Q4_K_M, PaddleOCR-VL-1.6 is the smaller download — 300,219,456 bytes against 397,807,936. At long context the gap widens: Qwen2.5-0.5B-Instruct's KV cache at 32K is 1.5× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	PaddleOCR-VL-1.6	Qwen2.5-0.5B-Instruct
Parameters	959M	494M
Architecture	paddleocr	qwen2
Layers	18	24
Native context	131,072	32,768
Mixture of experts	no	no
Quantizations published	36	32
Smallest quantization	0.15 GiB	0.31 GiB
Q4_K_M	0.28 GiB	0.37 GiB
Licence	apache-2.0	apache-2.0

KV cache by context

the term that decides long-context viability

Context	PaddleOCR-VL-1.6	Qwen2.5-0.5B-Instruct	Ratio
4,096	0.07 GiB	0.05 GiB	1.50×
8,192	0.14 GiB	0.09 GiB	1.50×
16,384	0.28 GiB	0.19 GiB	1.50×
32,768	0.56 GiB	0.38 GiB	1.50×
65,536	1.13 GiB	0.75 GiB	1.50×
131,072	2.25 GiB	1.50 GiB	1.50×