Skip to content

Surya OCR 2.5 VRAM requirements

How much VRAM Surya OCR 2.5 needs
4-bit weights: 399 MBWith 8K context: 1.2 GBWith 32K context: 1.4 GB8-bit, 8K context: 1.4 GB
Parameters
689M
Longest context
256K
Released
Oct 2026
License
OpenRAIL
Strongest at
Reading documents

How much VRAM Surya OCR 2.5 needs

Memory Surya OCR 2.5 needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit1.3 GB2.0 GB2.3 GB3.5 GB
8-bit699 MB1.4 GB1.7 GB2.9 GB
6-bit539 MB1.3 GB1.6 GB2.7 GB
5-bit468 MB1.2 GB1.5 GB2.6 GB
4-bit399 MB1.2 GB1.4 GB2.6 GB
3-bit321 MB1.1 GB1.4 GB2.5 GB
2-bit247 MB1.0 GB1.3 GB2.4 GB
1.6-bit136 MB916 MB1.2 GB2.3 GB

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

GPUs and Macs that run Surya OCR 2.5

The best way Surya OCR 2.5 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit256K214 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 4060 report
RTX 50608-bit256K321 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 5060 report
RTX 30708-bit256K321 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 3070 report
RTX 3080 10 GB8-bit256K469 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 3080 10 GB report
RTX 3060 12 GB8-bit256K270 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 3060 12 GB report
Arc B5808-bit256K211 tok/s0.1 sPDF: Surya OCR 2.5 on Arc B580 report
RTX 40708-bit256K351 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 4070 report
RTX 4070 Super8-bit256K351 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 4070 Super report
RTX 50708-bit256K431 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 5070 report
RTX 4060 Ti 16 GB8-bit256K225 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit256K321 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit256K242 tok/s0.1 sPDF: Surya OCR 2.5 on Arc A770 16 GB report
Radeon RX 9070 XT8-bit256K317 tok/s0.1 sPDF: Surya OCR 2.5 on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit256K431 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 4070 Ti Super report
RTX 4080 Super8-bit256K459 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 4080 Super report
RTX 5070 Ti8-bit256K521 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 5070 Ti report
RTX 50808-bit256K544 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 5080 report
Radeon RX 7900 XTX8-bit256K399 tok/s0.1 sPDF: Surya OCR 2.5 on Radeon RX 7900 XTX report
RTX 30908-bit256K535 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 3090 report
RTX 40908-bit256K560 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 4090 report
Radeon AI PRO R97008-bit256K317 tok/s0.1 sPDF: Surya OCR 2.5 on Radeon AI PRO R9700 report
RTX 50908-bit256K758 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 5090 report
2× RTX 30908-bit256K535 tok/s0.1 sPDF: Surya OCR 2.5 on 2× RTX 3090 report
2× RTX 40908-bit256K560 tok/s0.1 sPDF: Surya OCR 2.5 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit256K658 tok/s0.1 sPDF: Surya OCR 2.5 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit256K544 tok/s0.1 sPDF: Surya OCR 2.5 on RTX 6000 Ada report
L40S8-bit256K509 tok/s0.1 sPDF: Surya OCR 2.5 on L40S report
2× RTX 50908-bit256K758 tok/s0.1 sPDF: Surya OCR 2.5 on 2× RTX 5090 report
4× RTX 30908-bit256K535 tok/s0.1 sPDF: Surya OCR 2.5 on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit256K758 tok/s0.1 sPDF: Surya OCR 2.5 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit256K105 tok/s0.9 sPDF: Surya OCR 2.5 on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit256K105 tok/s0.9 sPDF: Surya OCR 2.5 on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit256K105 tok/s0.9 sPDF: Surya OCR 2.5 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit256K189 tok/s0.4 sPDF: Surya OCR 2.5 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit256K189 tok/s0.4 sPDF: Surya OCR 2.5 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit256K189 tok/s0.4 sPDF: Surya OCR 2.5 on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit256K266 tok/s0.2 sPDF: Surya OCR 2.5 on M4 Max 128 GB report
M5 Max 128 GB8-bit256K281 tok/s0.1 sPDF: Surya OCR 2.5 on M5 Max 128 GB report
M3 Ultra 96 GB8-bit256K300 tok/s0.1 sPDF: Surya OCR 2.5 on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit256K300 tok/s0.1 sPDF: Surya OCR 2.5 on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit256K300 tok/s0.1 sPDF: Surya OCR 2.5 on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit256K347 tok/s0.1 sPDF: Surya OCR 2.5 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit256K164 tok/s0.1 sPDF: Surya OCR 2.5 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit256K164 tok/s0.1 sPDF: Surya OCR 2.5 on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit256K233 tok/s0.1 sPDF: Surya OCR 2.5 on RTX Spark report
DGX Spark8-bit256K215 tok/s0.1 sPDF: Surya OCR 2.5 on DGX Spark report
Dual DGX Spark8-bit256K215 tok/s0.1 sPDF: Surya OCR 2.5 on Dual DGX Spark report
DGX Station GB3008-bit256K1,148 tok/s0.1 sPDF: Surya OCR 2.5 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit256K758 tok/s0.1 sPDF: Surya OCR 2.5 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit256K758 tok/s0.1 sPDF: Surya OCR 2.5 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit256K758 tok/s0.1 sPDF: Surya OCR 2.5 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit256K802 tok/s0.1 sPDF: Surya OCR 2.5 on A100 80 GB SXM report
H100 SXM8-bit256K961 tok/s0.1 sPDF: Surya OCR 2.5 on H100 SXM report
4× H200 NVL8-bit256K1,060 tok/s0.1 sPDF: Surya OCR 2.5 on 4× H200 NVL report
8× H200 NVL8-bit256K1,060 tok/s0.1 sPDF: Surya OCR 2.5 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 datalab-to/surya-ocr-2.5-gguf
vLLM
vllm serve datalab-to/surya-ocr-2.5

Questions

How much VRAM does Surya OCR 2.5 need?

Surya OCR 2.5 needs 399 MB for its weights at 4-bit, 1.2 GB with an 8K context and 1.4 GB with 32K. At 8-bit, 1.4 GB.

What is the minimum VRAM for Surya OCR 2.5?

About 1.1 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 Surya OCR 2.5?

689M parameters: 1.3 GB of weights at 16-bit, 399 MB at 4-bit.

Can Surya OCR 2.5 run on an RTX 4090?

Yes. At 8-bit it holds 256K of context and writes about 560 tokens a second.

More models