How much VRAM does PaddleOCR-VL-1.6 need?

Q4_K_M is exactly 300,219,456 bytes (0.28 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is PaddleOCR-VL-1.6's KV cache?

0.56 GiB at 32K context with an f16 cache, computed per layer from the model's architecture. Quantizing the cache to q8_0 roughly halves it.

PaddlePaddle · text

PaddleOCR-VL-1.6

PaddlePaddle/PaddleOCR-VL-1.6

PaddleOCR-VL-1.6 at Q4_K_M is exactly 300,219,456 bytes (0.28 GiB / 0.30 GB) — an effective 2.506 bits per weight, not the nominal 4. Its KV cache at 32K is 0.56 GiB.

From the file· summed from 1 file(s)From the file· KV per layer

Parameters

959M

Architecture

paddleocr

18 layers

Context

131,072

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
I1-IQ1_S	0.15 GiB	164,791,616	1.375	—	mradermacher
I1-IQ1_M	0.16 GiB	169,620,800	1.416	—	mradermacher
I1-IQ2_XXS	0.17 GiB	177,669,440	1.483	—	mradermacher
I1-IQ2_XS	0.17 GiB	185,288,000	1.546	—	mradermacher
I1-IQ2_S	0.19 GiB	201,991,488	1.686	—	mradermacher
I1-IQ2_M	0.19 GiB	208,430,400	1.740	—	mradermacher
I1-Q2_K_S	0.20 GiB	210,351,424	1.756	—	mradermacher
I1-IQ3_XXS	0.20 GiB	218,359,104	1.822	—	mradermacher
Q2_K	0.20 GiB	218,461,248	1.823	—	mradermacher
I1-Q2_K	0.20 GiB	218,461,504	1.823	—	mradermacher
I1-IQ3_XS	0.22 GiB	239,207,744	1.996	—	mradermacher
Q3_K_S	0.23 GiB	244,405,312	2.040	—	mradermacher
I1-Q3_K_S	0.23 GiB	244,405,568	2.040	—	mradermacher
I1-IQ3_S	0.23 GiB	245,032,256	2.045	—	mradermacher
I1-IQ3_M	0.23 GiB	250,881,344	2.094	—	mradermacher
Q3_K_M	0.24 GiB	258,024,512	2.153	—	mradermacher
I1-Q3_K_M	0.24 GiB	258,024,768	2.153	—	mradermacher
Q3_K_L	0.25 GiB	269,952,064	2.253	—	mradermacher
I1-Q3_K_L	0.25 GiB	269,952,320	2.253	—	mradermacher
I1-IQ4_XS	0.26 GiB	281,777,472	2.352	—	mradermacher
IQ4_XS	0.26 GiB	282,760,256	2.360	—	mradermacher
I1-Q4_0	0.27 GiB	292,705,600	2.443	—	mradermacher
I1-IQ4_NL	0.27 GiB	292,902,208	2.444	—	mradermacher
Q4_K_S	0.27 GiB	293,229,632	2.447	—	mradermacher
I1-Q4_K_S	0.27 GiB	293,229,888	2.447	—	mradermacher
Q4_K_M	0.28 GiB	300,219,456	2.506	—	mradermacher
I1-Q4_K_M	0.28 GiB	300,219,712	2.506	—	mradermacher
I1-Q4_1	0.29 GiB	314,856,768	2.628	—	mradermacher
Q5_K_S	0.31 GiB	337,400,896	2.816	—	mradermacher
I1-Q5_K_S	0.31 GiB	337,401,152	2.816	—	mradermacher
Q5_K_M	0.32 GiB	341,474,368	2.850	—	mradermacher
I1-Q5_K_M	0.32 GiB	341,474,624	2.850	—	mradermacher
Q6_K	0.36 GiB	385,307,712	3.216	—	mradermacher
I1-Q6_K	0.36 GiB	385,307,968	3.216	—	mradermacher
Q8_0	0.46 GiB	498,316,352	4.159	—	mradermacher
F16	0.87 GiB	935,769,152	7.810	—	mradermacher

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.07 GiB	0.07 GiB	—	18 / 0 / 0
8,192	0.14 GiB	0.14 GiB	—	18 / 0 / 0
16,384	0.28 GiB	0.28 GiB	—	18 / 0 / 0
32,768	0.56 GiB	0.56 GiB	—	18 / 0 / 0
65,536	1.13 GiB	1.13 GiB	—	18 / 0 / 0
131,072	2.25 GiB	2.25 GiB	—	18 / 0 / 0

Compare with

same modality, comparable size

vs ced-base vs embeddinggemma-300m vs Qwen3-0.6B vs Qwen2.5-0.5B-Instruct vs SmolLM2-135M-Instruct vs gemma-3-270m-it All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 0.50 GiB. The real file is 0.28 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

128

Hidden size

1024

Vocab

103,424

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does PaddleOCR-VL-1.6 need?: Q4_K_M is exactly 300,219,456 bytes (0.28 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is PaddleOCR-VL-1.6's KV cache?: 0.56 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Which quantization of PaddleOCR-VL-1.6 should I use?: Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.