Skip to content

HunyuanOCR VRAM requirements

How much VRAM HunyuanOCR needs
4-bit weights: 648 MBWith 8K context: 2.1 GBWith 32K context: 4.3 GB8-bit, 8K context: 2.6 GB
Parameters
1.1B
Longest context
128K
Released
Nov 2025
License
Tencent Hunyuan Community
Strongest at
Text and images

Its licence (Tencent Hunyuan Community) is not a standard open-source one: check the terms before commercial use.

How much VRAM HunyuanOCR needs

Memory HunyuanOCR needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit2.1 GB3.5 GB5.8 GB14.8 GB
8-bit1.1 GB2.6 GB4.8 GB13.8 GB
6-bit876 MB2.3 GB4.6 GB13.6 GB
5-bit760 MB2.2 GB4.5 GB13.5 GB
4-bit648 MB2.1 GB4.3 GB13.3 GB
3-bit522 MB2.0 GB4.2 GB13.2 GB
2-bit401 MB1.9 GB4.1 GB13.1 GB
1.6-bit220 MB1.7 GB3.9 GB12.9 GB

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

GPUs and Macs that run HunyuanOCR

The best way HunyuanOCR runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit60K143 tok/s0.1 sPDF: HunyuanOCR on RTX 4060 report
RTX 50608-bit60K222 tok/s0.1 sPDF: HunyuanOCR on RTX 5060 report
RTX 30708-bit60K222 tok/s0.1 sPDF: HunyuanOCR on RTX 3070 report
RTX 3080 10 GB8-bit80K338 tok/s0.1 sPDF: HunyuanOCR on RTX 3080 10 GB report
RTX 3060 12 GB8-bit96K184 tok/s0.2 sPDF: HunyuanOCR on RTX 3060 12 GB report
Arc B5808-bit96K150 tok/s0.1 sPDF: HunyuanOCR on Arc B580 report
RTX 40708-bit96K244 tok/s0.1 sPDF: HunyuanOCR on RTX 4070 report
RTX 4070 Super8-bit96K244 tok/s0.1 sPDF: HunyuanOCR on RTX 4070 Super report
RTX 50708-bit96K308 tok/s0.1 sPDF: HunyuanOCR on RTX 5070 report
RTX 4060 Ti 16 GB8-bit128K151 tok/s0.1 sPDF: HunyuanOCR on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit128K222 tok/s0.1 sPDF: HunyuanOCR on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit128K176 tok/s0.1 sPDF: HunyuanOCR on Arc A770 16 GB report
Radeon RX 9070 XT8-bit128K233 tok/s0.1 sPDF: HunyuanOCR on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit128K308 tok/s0.1 sPDF: HunyuanOCR on RTX 4070 Ti Super report
RTX 4080 Super8-bit128K330 tok/s0.1 sPDF: HunyuanOCR on RTX 4080 Super report
RTX 5070 Ti8-bit128K382 tok/s0.1 sPDF: HunyuanOCR on RTX 5070 Ti report
RTX 50808-bit128K401 tok/s0.1 sPDF: HunyuanOCR on RTX 5080 report
Radeon RX 7900 XTX8-bit128K306 tok/s0.1 sPDF: HunyuanOCR on Radeon RX 7900 XTX report
RTX 30908-bit128K394 tok/s0.1 sPDF: HunyuanOCR on RTX 3090 report
RTX 40908-bit128K416 tok/s0.1 sPDF: HunyuanOCR on RTX 4090 report
Radeon AI PRO R97008-bit128K233 tok/s0.1 sPDF: HunyuanOCR on Radeon AI PRO R9700 report
RTX 50908-bit128K599 tok/s0.1 sPDF: HunyuanOCR on RTX 5090 report
2× RTX 30908-bit128K394 tok/s0.1 sPDF: HunyuanOCR on 2× RTX 3090 report
2× RTX 40908-bit128K416 tok/s0.1 sPDF: HunyuanOCR on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit128K504 tok/s0.1 sPDF: HunyuanOCR on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit128K401 tok/s0.1 sPDF: HunyuanOCR on RTX 6000 Ada report
L40S8-bit128K372 tok/s0.1 sPDF: HunyuanOCR on L40S report
2× RTX 50908-bit128K599 tok/s0.1 sPDF: HunyuanOCR on 2× RTX 5090 report
4× RTX 30908-bit128K394 tok/s0.1 sPDF: HunyuanOCR on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit128K599 tok/s0.1 sPDF: HunyuanOCR on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit88K71 tok/s1.4 sPDF: HunyuanOCR on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit128K71 tok/s1.4 sPDF: HunyuanOCR on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit128K71 tok/s1.4 sPDF: HunyuanOCR on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit128K138 tok/s0.7 sPDF: HunyuanOCR on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit128K138 tok/s0.7 sPDF: HunyuanOCR on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit128K138 tok/s0.7 sPDF: HunyuanOCR on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit128K207 tok/s0.4 sPDF: HunyuanOCR on M4 Max 128 GB report
M5 Max 128 GB8-bit128K222 tok/s0.1 sPDF: HunyuanOCR on M5 Max 128 GB report
M3 Ultra 96 GB8-bit128K241 tok/s0.2 sPDF: HunyuanOCR on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit128K241 tok/s0.2 sPDF: HunyuanOCR on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit128K241 tok/s0.2 sPDF: HunyuanOCR on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit128K290 tok/s0.1 sPDF: HunyuanOCR on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit128K112 tok/s0.2 sPDF: HunyuanOCR on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit128K112 tok/s0.2 sPDF: HunyuanOCR on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit128K157 tok/s0.1 sPDF: HunyuanOCR on RTX Spark report
DGX Spark8-bit128K144 tok/s0.1 sPDF: HunyuanOCR on DGX Spark report
Dual DGX Spark8-bit128K144 tok/s0.1 sPDF: HunyuanOCR on Dual DGX Spark report
DGX Station GB3008-bit128K1,042 tok/s0.1 sPDF: HunyuanOCR on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit128K599 tok/s0.1 sPDF: HunyuanOCR on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit128K599 tok/s0.1 sPDF: HunyuanOCR on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit128K599 tok/s0.1 sPDF: HunyuanOCR on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit128K644 tok/s0.1 sPDF: HunyuanOCR on A100 80 GB SXM report
H100 SXM8-bit128K815 tok/s0.1 sPDF: HunyuanOCR on H100 SXM report
4× H200 NVL8-bit128K931 tok/s0.1 sPDF: HunyuanOCR on 4× H200 NVL report
8× H200 NVL8-bit128K931 tok/s0.1 sPDF: HunyuanOCR 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 prithivMLmods/HunyuanOCR-1.5-GGUF-Updated
vLLM
vllm serve tencent/HunyuanOCR

Questions

How much VRAM does HunyuanOCR need?

HunyuanOCR needs 648 MB for its weights at 4-bit, 2.1 GB with an 8K context and 4.3 GB with 32K. At 8-bit, 2.6 GB.

What is the minimum VRAM for HunyuanOCR?

About 2.0 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 HunyuanOCR?

1.1B parameters: 2.1 GB of weights at 16-bit, 648 MB at 4-bit.

Can HunyuanOCR run on an RTX 4090?

Yes. At 8-bit it holds 128K of context and writes about 416 tokens a second.

More models