Skip to content

PaddleOCR VL 1.6 VRAM requirements

How much VRAM PaddleOCR VL 1.6 needs
4-bit weights: 554 MBWith 8K context: 1.4 GBWith 32K context: 1.8 GB8-bit, 8K context: 1.8 GB
Parameters
959M
Longest context
128K
Released
May 2026
License
Apache 2.0
Strongest at
Text and images

How much VRAM PaddleOCR VL 1.6 needs

Memory PaddleOCR VL 1.6 needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit1.8 GB2.6 GB3.0 GB4.7 GB
8-bit971 MB1.8 GB2.2 GB3.9 GB
6-bit750 MB1.6 GB2.0 GB3.7 GB
5-bit650 MB1.5 GB1.9 GB3.6 GB
4-bit554 MB1.4 GB1.8 GB3.5 GB
3-bit447 MB1.3 GB1.7 GB3.4 GB
2-bit343 MB1.2 GB1.6 GB3.3 GB
1.6-bit189 MB1.0 GB1.4 GB3.1 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run PaddleOCR VL 1.6

The best way PaddleOCR VL 1.6 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit128K189 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 4060 report
RTX 50608-bit128K292 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 5060 report
RTX 30708-bit128K292 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 3070 report
RTX 3080 10 GB8-bit128K446 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 3080 10 GB report
RTX 3060 12 GB8-bit128K242 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 3060 12 GB report
Arc B5808-bit128K198 tok/s0.1 sPDF: PaddleOCR VL 1.6 on Arc B580 report
RTX 40708-bit128K322 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 4070 report
RTX 4070 Super8-bit128K322 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 4070 Super report
RTX 50708-bit128K406 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 5070 report
RTX 4060 Ti 16 GB8-bit128K199 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit128K292 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit128K232 tok/s0.1 sPDF: PaddleOCR VL 1.6 on Arc A770 16 GB report
Radeon RX 9070 XT8-bit128K307 tok/s0.1 sPDF: PaddleOCR VL 1.6 on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit128K406 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 4070 Ti Super report
RTX 4080 Super8-bit128K435 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 4080 Super report
RTX 5070 Ti8-bit128K504 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 5070 Ti report
RTX 50808-bit128K530 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 5080 report
Radeon RX 7900 XTX8-bit128K404 tok/s0.1 sPDF: PaddleOCR VL 1.6 on Radeon RX 7900 XTX report
RTX 30908-bit128K520 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 3090 report
RTX 40908-bit128K548 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 4090 report
Radeon AI PRO R97008-bit128K307 tok/s0.1 sPDF: PaddleOCR VL 1.6 on Radeon AI PRO R9700 report
RTX 50908-bit128K792 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 5090 report
2× RTX 30908-bit128K520 tok/s0.1 sPDF: PaddleOCR VL 1.6 on 2× RTX 3090 report
2× RTX 40908-bit128K548 tok/s0.1 sPDF: PaddleOCR VL 1.6 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit128K665 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit128K530 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX 6000 Ada report
L40S8-bit128K491 tok/s0.1 sPDF: PaddleOCR VL 1.6 on L40S report
2× RTX 50908-bit128K792 tok/s0.1 sPDF: PaddleOCR VL 1.6 on 2× RTX 5090 report
4× RTX 30908-bit128K520 tok/s0.1 sPDF: PaddleOCR VL 1.6 on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit128K792 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit128K94 tok/s1.1 sPDF: PaddleOCR VL 1.6 on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit128K94 tok/s1.1 sPDF: PaddleOCR VL 1.6 on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit128K94 tok/s1.1 sPDF: PaddleOCR VL 1.6 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit128K182 tok/s0.5 sPDF: PaddleOCR VL 1.6 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit128K182 tok/s0.5 sPDF: PaddleOCR VL 1.6 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit128K182 tok/s0.5 sPDF: PaddleOCR VL 1.6 on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit128K274 tok/s0.3 sPDF: PaddleOCR VL 1.6 on M4 Max 128 GB report
M5 Max 128 GB8-bit128K293 tok/s0.1 sPDF: PaddleOCR VL 1.6 on M5 Max 128 GB report
M3 Ultra 96 GB8-bit128K318 tok/s0.2 sPDF: PaddleOCR VL 1.6 on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit128K318 tok/s0.2 sPDF: PaddleOCR VL 1.6 on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit128K318 tok/s0.2 sPDF: PaddleOCR VL 1.6 on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit128K384 tok/s0.1 sPDF: PaddleOCR VL 1.6 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit128K147 tok/s0.2 sPDF: PaddleOCR VL 1.6 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit128K147 tok/s0.2 sPDF: PaddleOCR VL 1.6 on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit128K206 tok/s0.1 sPDF: PaddleOCR VL 1.6 on RTX Spark report
DGX Spark8-bit128K189 tok/s0.1 sPDF: PaddleOCR VL 1.6 on DGX Spark report
Dual DGX Spark8-bit128K189 tok/s0.1 sPDF: PaddleOCR VL 1.6 on Dual DGX Spark report
DGX Station GB3008-bit128K1,384 tok/s0.1 sPDF: PaddleOCR VL 1.6 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit128K792 tok/s0.1 sPDF: PaddleOCR VL 1.6 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit128K792 tok/s0.1 sPDF: PaddleOCR VL 1.6 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit128K792 tok/s0.1 sPDF: PaddleOCR VL 1.6 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit128K851 tok/s0.1 sPDF: PaddleOCR VL 1.6 on A100 80 GB SXM report
H100 SXM8-bit128K1,080 tok/s0.1 sPDF: PaddleOCR VL 1.6 on H100 SXM report
4× H200 NVL8-bit128K1,235 tok/s0.1 sPDF: PaddleOCR VL 1.6 on 4× H200 NVL report
8× H200 NVL8-bit128K1,235 tok/s0.1 sPDF: PaddleOCR VL 1.6 on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

llama.cpp
llama-server -hf LunarOilRig/PaddleOCR-VL-1.6-GGUF-Q4
vLLM
vllm serve PaddlePaddle/PaddleOCR-VL-1.6
MLX (Mac)
mlx_lm.chat --model mlx-community/PaddleOCR-VL-1.6-4bit

Questions

How much VRAM does PaddleOCR VL 1.6 need?

PaddleOCR VL 1.6 needs 554 MB for its weights at 4-bit, 1.4 GB with an 8K context and 1.8 GB with 32K. At 8-bit, 1.8 GB.

What is the minimum VRAM for PaddleOCR VL 1.6?

About 1.3 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the RTX 4060.

How big is PaddleOCR VL 1.6?

959M parameters: 1.8 GB of weights at 16-bit, 554 MB at 4-bit.

Can PaddleOCR VL 1.6 run on an RTX 4090?

Yes. At 8-bit it holds 128K of context and writes about 548 tokens a second.

More models