Skip to content

DeepSeek OCR 2 VRAM requirements

How much VRAM DeepSeek OCR 2 needs
4-bit weights: 1.9 GBWith 8K context: 3.1 GB8-bit, 8K context: 4.6 GB
Parameters
3.4B
Active per token
1.2B
Longest context
8K
Released
Jan 2026
License
Apache 2.0
Strongest at
Reading documents

How much VRAM DeepSeek OCR 2 needs

Memory DeepSeek OCR 2 needs at each precision and context length
PrecisionWeightsWith 8K
16-bit6.3 GB7.5 GB
8-bit3.4 GB4.6 GB
6-bit2.6 GB3.8 GB
5-bit2.2 GB3.4 GB
4-bit1.9 GB3.1 GB
3-bit1.5 GB2.7 GB
2-bit1.2 GB2.4 GB
1.6-bit667 MB1.9 GB

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

GPUs and Macs that run DeepSeek OCR 2

The best way DeepSeek OCR 2 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit8K135 tok/s0.4 sPDF: DeepSeek OCR 2 on RTX 4060 report
RTX 50608-bit8K210 tok/s0.3 sPDF: DeepSeek OCR 2 on RTX 5060 report
RTX 30708-bit8K210 tok/s0.3 sPDF: DeepSeek OCR 2 on RTX 3070 report
RTX 3080 10 GB8-bit8K322 tok/s0.2 sPDF: DeepSeek OCR 2 on RTX 3080 10 GB report
RTX 3060 12 GB8-bit8K174 tok/s0.5 sPDF: DeepSeek OCR 2 on RTX 3060 12 GB report
Arc B5808-bit8K148 tok/s0.3 sPDF: DeepSeek OCR 2 on Arc B580 report
RTX 40708-bit8K232 tok/s0.2 sPDF: DeepSeek OCR 2 on RTX 4070 report
RTX 4070 Super8-bit8K232 tok/s0.2 sPDF: DeepSeek OCR 2 on RTX 4070 Super report
RTX 50708-bit8K293 tok/s0.2 sPDF: DeepSeek OCR 2 on RTX 5070 report
RTX 4060 Ti 16 GB8-bit8K143 tok/s0.3 sPDF: DeepSeek OCR 2 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit8K210 tok/s0.3 sPDF: DeepSeek OCR 2 on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit8K174 tok/s0.3 sPDF: DeepSeek OCR 2 on Arc A770 16 GB report
Radeon RX 9070 XT8-bit8K228 tok/s0.2 sPDF: DeepSeek OCR 2 on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit8K293 tok/s0.1 sPDF: DeepSeek OCR 2 on RTX 4070 Ti Super report
RTX 4080 Super8-bit8K314 tok/s0.1 sPDF: DeepSeek OCR 2 on RTX 4080 Super report
RTX 5070 Ti8-bit8K365 tok/s0.1 sPDF: DeepSeek OCR 2 on RTX 5070 Ti report
RTX 50808-bit8K383 tok/s0.1 sPDF: DeepSeek OCR 2 on RTX 5080 report
Radeon RX 7900 XTX8-bit8K305 tok/s0.3 sPDF: DeepSeek OCR 2 on Radeon RX 7900 XTX report
RTX 30908-bit8K376 tok/s0.2 sPDF: DeepSeek OCR 2 on RTX 3090 report
RTX 40908-bit8K397 tok/s0.1 sPDF: DeepSeek OCR 2 on RTX 4090 report
Radeon AI PRO R97008-bit8K228 tok/s0.2 sPDF: DeepSeek OCR 2 on Radeon AI PRO R9700 report
RTX 50908-bit8K578 tok/s0.1 sPDF: DeepSeek OCR 2 on RTX 5090 report
2× RTX 30908-bit8K376 tok/s0.1 sPDF: DeepSeek OCR 2 on 2× RTX 3090 report
2× RTX 40908-bit8K397 tok/s0.1 sPDF: DeepSeek OCR 2 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit8K483 tok/s0.1 sPDF: DeepSeek OCR 2 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit8K383 tok/s0.1 sPDF: DeepSeek OCR 2 on RTX 6000 Ada report
L40S8-bit8K355 tok/s0.1 sPDF: DeepSeek OCR 2 on L40S report
2× RTX 50908-bit8K578 tok/s0.1 sPDF: DeepSeek OCR 2 on 2× RTX 5090 report
4× RTX 30908-bit8K376 tok/s0.1 sPDF: DeepSeek OCR 2 on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit8K578 tok/s0.1 sPDF: DeepSeek OCR 2 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit8K69 tok/s4.1 sPDF: DeepSeek OCR 2 on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit8K69 tok/s4.1 sPDF: DeepSeek OCR 2 on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit8K69 tok/s4.1 sPDF: DeepSeek OCR 2 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit8K136 tok/s2.0 sPDF: DeepSeek OCR 2 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit8K136 tok/s2.0 sPDF: DeepSeek OCR 2 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit8K136 tok/s2.0 sPDF: DeepSeek OCR 2 on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit8K209 tok/s1.0 sPDF: DeepSeek OCR 2 on M4 Max 128 GB report
M5 Max 128 GB8-bit8K225 tok/s0.3 sPDF: DeepSeek OCR 2 on M5 Max 128 GB report
M3 Ultra 96 GB8-bit8K246 tok/s0.6 sPDF: DeepSeek OCR 2 on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit8K246 tok/s0.6 sPDF: DeepSeek OCR 2 on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit8K246 tok/s0.6 sPDF: DeepSeek OCR 2 on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit8K302 tok/s0.2 sPDF: DeepSeek OCR 2 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit8K107 tok/s0.6 sPDF: DeepSeek OCR 2 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit8K107 tok/s0.6 sPDF: DeepSeek OCR 2 on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit8K148 tok/s0.2 sPDF: DeepSeek OCR 2 on RTX Spark report
DGX Spark8-bit8K136 tok/s0.2 sPDF: DeepSeek OCR 2 on DGX Spark report
Dual DGX Spark8-bit8K136 tok/s0.1 sPDF: DeepSeek OCR 2 on Dual DGX Spark report
DGX Station GB3008-bit8K1,025 tok/s0.1 sPDF: DeepSeek OCR 2 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit8K578 tok/s0.1 sPDF: DeepSeek OCR 2 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit8K578 tok/s0.1 sPDF: DeepSeek OCR 2 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit8K578 tok/s0.1 sPDF: DeepSeek OCR 2 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit8K622 tok/s0.1 sPDF: DeepSeek OCR 2 on A100 80 GB SXM report
H100 SXM8-bit8K793 tok/s0.1 sPDF: DeepSeek OCR 2 on H100 SXM report
4× H200 NVL8-bit8K911 tok/s0.1 sPDF: DeepSeek OCR 2 on 4× H200 NVL report
8× H200 NVL8-bit8K911 tok/s0.1 sPDF: DeepSeek OCR 2 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 sabafallah/DeepSeek-OCR-2-GGUF
vLLM
vllm serve deepseek-ai/DeepSeek-OCR-2

Questions

How much VRAM does DeepSeek OCR 2 need?

DeepSeek OCR 2 needs 1.9 GB for its weights at 4-bit, 3.1 GB with an 8K context and 3.1 GB with 8K. At 8-bit, 4.6 GB.

What is the minimum VRAM for DeepSeek OCR 2?

About 2.7 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 DeepSeek OCR 2?

3.4B parameters (1.2B active per token): 6.3 GB of weights at 16-bit, 1.9 GB at 4-bit.

Can DeepSeek OCR 2 run on an RTX 4090?

Yes. At 8-bit it holds 8K of context and writes about 397 tokens a second.

More models