Is PaddleOCR-VL-1.6 or embeddinggemma-300m smaller?

Both publish quantizations; compare the table for the exact sizes.

Which uses less memory at long context?

embeddinggemma-300m — its KV cache at 32K context is 3.9x smaller, because of how its attention layers cache.

Model comparison

PaddleOCR-VL-1.6 vs embeddinggemma-300m

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: embeddinggemma-300m's KV cache at 32K is 3.9× 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	embeddinggemma-300m
Parameters	959M	303M
Architecture	paddleocr	gemma-embedding
Layers	18	24
Native context	131,072	2,048
Mixture of experts	no	no
Quantizations published	36	10
Smallest quantization	0.15 GiB	0.26 GiB
Q4_K_M	0.28 GiB	—
Licence	apache-2.0	—

KV cache by context

the term that decides long-context viability

Context	PaddleOCR-VL-1.6	embeddinggemma-300m	Ratio
4,096	0.07 GiB	0.04 GiB	2.00×
8,192	0.14 GiB	0.05 GiB	2.77×
16,384	0.28 GiB	0.08 GiB	3.43×
32,768	0.56 GiB	0.14 GiB	3.89×
65,536	1.13 GiB	0.27 GiB	4.17×
131,072	2.25 GiB	0.52 GiB	4.33×