Skip to content

Qwen2.5 VL 7B Instruct VRAM requirements

How much VRAM Qwen2.5 VL 7B Instruct needs
4-bit weights: 4.7 GBWith 8K context: 6.0 GBWith 32K context: 7.3 GB8-bit, 8K context: 9.5 GB
Parameters
8.3B
Longest context
125K
Released
Jan 2025
License
Apache 2.0
Strongest at
Text and images

How much VRAM Qwen2.5 VL 7B Instruct needs

Memory Qwen2.5 VL 7B Instruct needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 125K
16-bit15.4 GB16.7 GB18.0 GB23.1 GB
8-bit8.2 GB9.5 GB10.8 GB15.9 GB
6-bit6.3 GB7.6 GB8.9 GB14.0 GB
5-bit5.5 GB6.8 GB8.1 GB13.2 GB
4-bit4.7 GB6.0 GB7.3 GB12.4 GB
3-bit3.8 GB5.1 GB6.4 GB11.5 GB
2-bit2.9 GB4.2 GB5.5 GB10.6 GB
1.6-bit1.6 GB2.9 GB4.2 GB9.3 GB

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

GPUs and Macs that run Qwen2.5 VL 7B Instruct

The best way Qwen2.5 VL 7B Instruct runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40605-bit20K34 tok/s1.1 sPDF: Qwen2.5 VL 7B Instruct on RTX 4060 report
RTX 50605-bit20K54 tok/s0.8 sPDF: Qwen2.5 VL 7B Instruct on RTX 5060 report
RTX 30705-bit20K54 tok/s0.8 sPDF: Qwen2.5 VL 7B Instruct on RTX 3070 report
RTX 3080 10 GB6-bit40K78 tok/s0.5 sPDF: Qwen2.5 VL 7B Instruct on RTX 3080 10 GB report
RTX 3060 12 GB8-bit40K30 tok/s1.3 sPDF: Qwen2.5 VL 7B Instruct on RTX 3060 12 GB report
Arc B5808-bit40K26 tok/s0.9 sPDF: Qwen2.5 VL 7B Instruct on Arc B580 report
RTX 40708-bit40K41 tok/s0.5 sPDF: Qwen2.5 VL 7B Instruct on RTX 4070 report
RTX 4070 Super8-bit40K41 tok/s0.4 sPDF: Qwen2.5 VL 7B Instruct on RTX 4070 Super report
RTX 50708-bit40K55 tok/s0.5 sPDF: Qwen2.5 VL 7B Instruct on RTX 5070 report
RTX 4060 Ti 16 GB8-bit112K24 tok/s0.7 sPDF: Qwen2.5 VL 7B Instruct on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit112K37 tok/s0.7 sPDF: Qwen2.5 VL 7B Instruct on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit112K32 tok/s0.8 sPDF: Qwen2.5 VL 7B Instruct on Arc A770 16 GB report
Radeon RX 9070 XT8-bit112K44 tok/s0.4 sPDF: Qwen2.5 VL 7B Instruct on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit112K55 tok/s0.4 sPDF: Qwen2.5 VL 7B Instruct on RTX 4070 Ti Super report
RTX 4080 Super8-bit112K59 tok/s0.3 sPDF: Qwen2.5 VL 7B Instruct on RTX 4080 Super report
RTX 5070 Ti8-bit112K72 tok/s0.4 sPDF: Qwen2.5 VL 7B Instruct on RTX 5070 Ti report
RTX 50808-bit112K76 tok/s0.3 sPDF: Qwen2.5 VL 7B Instruct on RTX 5080 report
Radeon RX 7900 XTX8-bit125K64 tok/s0.7 sPDF: Qwen2.5 VL 7B Instruct on Radeon RX 7900 XTX report
RTX 30908-bit125K75 tok/s0.4 sPDF: Qwen2.5 VL 7B Instruct on RTX 3090 report
RTX 40908-bit125K80 tok/s0.2 sPDF: Qwen2.5 VL 7B Instruct on RTX 4090 report
Radeon AI PRO R97008-bit125K44 tok/s0.5 sPDF: Qwen2.5 VL 7B Instruct on Radeon AI PRO R9700 report
RTX 50908-bit125K135 tok/s0.2 sPDF: Qwen2.5 VL 7B Instruct on RTX 5090 report
2× RTX 30908-bit125K75 tok/s0.3 sPDF: Qwen2.5 VL 7B Instruct on 2× RTX 3090 report
2× RTX 40908-bit125K80 tok/s0.1 sPDF: Qwen2.5 VL 7B Instruct on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit125K104 tok/s0.2 sPDF: Qwen2.5 VL 7B Instruct on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit125K76 tok/s0.2 sPDF: Qwen2.5 VL 7B Instruct on RTX 6000 Ada report
L40S8-bit125K69 tok/s0.3 sPDF: Qwen2.5 VL 7B Instruct on L40S report
2× RTX 50908-bit125K135 tok/s0.1 sPDF: Qwen2.5 VL 7B Instruct on 2× RTX 5090 report
4× RTX 30908-bit125K75 tok/s0.1 sPDF: Qwen2.5 VL 7B Instruct on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit125K135 tok/s0.1 sPDF: Qwen2.5 VL 7B Instruct on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit20K12 tok/s10 sPDF: Qwen2.5 VL 7B Instruct on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit112K12 tok/s10 sPDF: Qwen2.5 VL 7B Instruct on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit125K12 tok/s10 sPDF: Qwen2.5 VL 7B Instruct on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit112K26 tok/s5.1 sPDF: Qwen2.5 VL 7B Instruct on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit125K26 tok/s5.1 sPDF: Qwen2.5 VL 7B Instruct on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit125K26 tok/s5.1 sPDF: Qwen2.5 VL 7B Instruct on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit125K45 tok/s2.5 sPDF: Qwen2.5 VL 7B Instruct on M4 Max 128 GB report
M5 Max 128 GB8-bit125K50 tok/s0.7 sPDF: Qwen2.5 VL 7B Instruct on M5 Max 128 GB report
M3 Ultra 96 GB8-bit125K56 tok/s1.4 sPDF: Qwen2.5 VL 7B Instruct on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit125K56 tok/s1.4 sPDF: Qwen2.5 VL 7B Instruct on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit125K56 tok/s1.4 sPDF: Qwen2.5 VL 7B Instruct on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit125K78 tok/s0.4 sPDF: Qwen2.5 VL 7B Instruct on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit125K18 tok/s1.5 sPDF: Qwen2.5 VL 7B Instruct on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit125K18 tok/s1.5 sPDF: Qwen2.5 VL 7B Instruct on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit125K25 tok/s0.5 sPDF: Qwen2.5 VL 7B Instruct on RTX Spark report
DGX Spark8-bit125K23 tok/s0.5 sPDF: Qwen2.5 VL 7B Instruct on DGX Spark report
Dual DGX Spark8-bit125K23 tok/s0.3 sPDF: Qwen2.5 VL 7B Instruct on Dual DGX Spark report
DGX Station GB3008-bit125K401 tok/s0.1 sPDF: Qwen2.5 VL 7B Instruct on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit125K135 tok/s0.1 sPDF: Qwen2.5 VL 7B Instruct on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit125K135 tok/s0.1 sPDF: Qwen2.5 VL 7B Instruct on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit125K135 tok/s0.1 sPDF: Qwen2.5 VL 7B Instruct on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit125K151 tok/s0.3 sPDF: Qwen2.5 VL 7B Instruct on A100 80 GB SXM report
H100 SXM8-bit125K230 tok/s0.1 sPDF: Qwen2.5 VL 7B Instruct on H100 SXM report
4× H200 NVL8-bit125K304 tok/s0.1 sPDF: Qwen2.5 VL 7B Instruct on 4× H200 NVL report
8× H200 NVL8-bit125K304 tok/s0.1 sPDF: Qwen2.5 VL 7B Instruct 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 unsloth/Qwen2.5-VL-7B-Instruct-GGUF
vLLM
vllm serve Qwen/Qwen2.5-VL-7B-Instruct
MLX (Mac)
mlx_lm.chat --model mlx-community/olmOCR-2-7B-1025-mlx-8bit

Questions

How much VRAM does Qwen2.5 VL 7B Instruct need?

Qwen2.5 VL 7B Instruct needs 4.7 GB for its weights at 4-bit, 6.0 GB with an 8K context and 7.3 GB with 32K. At 8-bit, 9.5 GB.

What is the minimum VRAM for Qwen2.5 VL 7B Instruct?

About 5.1 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 Qwen2.5 VL 7B Instruct?

8.3B parameters: 15.4 GB of weights at 16-bit, 4.7 GB at 4-bit.

Can Qwen2.5 VL 7B Instruct run on an RTX 4090?

Yes. At 8-bit it holds 125K of context and writes about 80 tokens a second.

More models