Skip to content

InternVL3 5 4B HF VRAM requirements

How much VRAM InternVL3 5 4B HF needs
4-bit weights: 2.7 GBWith 8K context: 4.6 GBWith 32K context: 8.0 GB8-bit, 8K context: 6.6 GB
Parameters
4.7B
Longest context
40K
Released
Aug 2025
License
Apache 2.0
Strongest at
Text and images

How much VRAM InternVL3 5 4B HF needs

Memory InternVL3 5 4B HF needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 40K
16-bit8.8 GB10.8 GB14.1 GB15.3 GB
8-bit4.7 GB6.6 GB10.0 GB11.1 GB
6-bit3.6 GB5.6 GB8.9 GB10.1 GB
5-bit3.1 GB5.1 GB8.5 GB9.6 GB
4-bit2.7 GB4.6 GB8.0 GB9.1 GB
3-bit2.2 GB4.1 GB7.5 GB8.6 GB
2-bit1.7 GB3.6 GB7.0 GB8.1 GB
1.6-bit931 MB2.9 GB6.2 GB7.4 GB

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

GPUs and Macs that run InternVL3 5 4B HF

The best way InternVL3 5 4B HF runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40606-bit20K51 tok/s0.6 sPDF: InternVL3 5 4B HF on RTX 4060 report
RTX 50606-bit20K80 tok/s0.5 sPDF: InternVL3 5 4B HF on RTX 5060 report
RTX 30706-bit20K80 tok/s0.4 sPDF: InternVL3 5 4B HF on RTX 3070 report
RTX 3080 10 GB8-bit28K103 tok/s0.3 sPDF: InternVL3 5 4B HF on RTX 3080 10 GB report
RTX 3060 12 GB8-bit40K52 tok/s0.7 sPDF: InternVL3 5 4B HF on RTX 3060 12 GB report
Arc B5808-bit40K45 tok/s0.5 sPDF: InternVL3 5 4B HF on Arc B580 report
RTX 40708-bit40K71 tok/s0.3 sPDF: InternVL3 5 4B HF on RTX 4070 report
RTX 4070 Super8-bit40K71 tok/s0.3 sPDF: InternVL3 5 4B HF on RTX 4070 Super report
RTX 50708-bit40K92 tok/s0.3 sPDF: InternVL3 5 4B HF on RTX 5070 report
RTX 4060 Ti 16 GB8-bit40K42 tok/s0.4 sPDF: InternVL3 5 4B HF on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit40K63 tok/s0.4 sPDF: InternVL3 5 4B HF on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit40K54 tok/s0.4 sPDF: InternVL3 5 4B HF on Arc A770 16 GB report
Radeon RX 9070 XT8-bit40K72 tok/s0.2 sPDF: InternVL3 5 4B HF on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit40K92 tok/s0.2 sPDF: InternVL3 5 4B HF on RTX 4070 Ti Super report
RTX 4080 Super8-bit40K100 tok/s0.2 sPDF: InternVL3 5 4B HF on RTX 4080 Super report
RTX 5070 Ti8-bit40K119 tok/s0.2 sPDF: InternVL3 5 4B HF on RTX 5070 Ti report
RTX 50808-bit40K126 tok/s0.2 sPDF: InternVL3 5 4B HF on RTX 5080 report
Radeon RX 7900 XTX8-bit40K102 tok/s0.4 sPDF: InternVL3 5 4B HF on Radeon RX 7900 XTX report
RTX 30908-bit40K123 tok/s0.3 sPDF: InternVL3 5 4B HF on RTX 3090 report
RTX 40908-bit40K131 tok/s0.1 sPDF: InternVL3 5 4B HF on RTX 4090 report
Radeon AI PRO R97008-bit40K72 tok/s0.3 sPDF: InternVL3 5 4B HF on Radeon AI PRO R9700 report
RTX 50908-bit40K210 tok/s0.1 sPDF: InternVL3 5 4B HF on RTX 5090 report
2× RTX 30908-bit40K123 tok/s0.1 sPDF: InternVL3 5 4B HF on 2× RTX 3090 report
2× RTX 40908-bit40K131 tok/s0.1 sPDF: InternVL3 5 4B HF on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit40K167 tok/s0.1 sPDF: InternVL3 5 4B HF on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit40K126 tok/s0.1 sPDF: InternVL3 5 4B HF on RTX 6000 Ada report
L40S8-bit40K115 tok/s0.2 sPDF: InternVL3 5 4B HF on L40S report
2× RTX 50908-bit40K210 tok/s0.1 sPDF: InternVL3 5 4B HF on 2× RTX 5090 report
4× RTX 30908-bit40K123 tok/s0.1 sPDF: InternVL3 5 4B HF on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit40K210 tok/s0.1 sPDF: InternVL3 5 4B HF on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit32K20 tok/s5.7 sPDF: InternVL3 5 4B HF on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit40K20 tok/s5.7 sPDF: InternVL3 5 4B HF on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit40K20 tok/s5.7 sPDF: InternVL3 5 4B HF on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit40K43 tok/s2.8 sPDF: InternVL3 5 4B HF on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit40K43 tok/s2.8 sPDF: InternVL3 5 4B HF on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit40K43 tok/s2.8 sPDF: InternVL3 5 4B HF on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit40K71 tok/s1.4 sPDF: InternVL3 5 4B HF on M4 Max 128 GB report
M5 Max 128 GB8-bit40K78 tok/s0.4 sPDF: InternVL3 5 4B HF on M5 Max 128 GB report
M3 Ultra 96 GB8-bit40K87 tok/s0.8 sPDF: InternVL3 5 4B HF on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit40K87 tok/s0.8 sPDF: InternVL3 5 4B HF on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit40K87 tok/s0.8 sPDF: InternVL3 5 4B HF on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit40K114 tok/s0.2 sPDF: InternVL3 5 4B HF on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit40K31 tok/s0.8 sPDF: InternVL3 5 4B HF on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit40K31 tok/s0.8 sPDF: InternVL3 5 4B HF on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit40K43 tok/s0.3 sPDF: InternVL3 5 4B HF on RTX Spark report
DGX Spark8-bit40K40 tok/s0.3 sPDF: InternVL3 5 4B HF on DGX Spark report
Dual DGX Spark8-bit40K40 tok/s0.2 sPDF: InternVL3 5 4B HF on Dual DGX Spark report
DGX Station GB3008-bit40K498 tok/s0.1 sPDF: InternVL3 5 4B HF on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit40K210 tok/s0.1 sPDF: InternVL3 5 4B HF on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit40K210 tok/s0.1 sPDF: InternVL3 5 4B HF on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit40K210 tok/s0.1 sPDF: InternVL3 5 4B HF on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit40K232 tok/s0.2 sPDF: InternVL3 5 4B HF on A100 80 GB SXM report
H100 SXM8-bit40K328 tok/s0.1 sPDF: InternVL3 5 4B HF on H100 SXM report
4× H200 NVL8-bit40K408 tok/s0.1 sPDF: InternVL3 5 4B HF on 4× H200 NVL report
8× H200 NVL8-bit40K408 tok/s0.1 sPDF: InternVL3 5 4B HF 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 lewiswatson/Frame2KG-InternVL3.5-4b-JSON
vLLM
vllm serve OpenGVLab/InternVL3_5-4B-HF

Questions

How much VRAM does InternVL3 5 4B HF need?

InternVL3 5 4B HF needs 2.7 GB for its weights at 4-bit, 4.6 GB with an 8K context and 8.0 GB with 32K. At 8-bit, 6.6 GB.

What is the minimum VRAM for InternVL3 5 4B HF?

About 4.1 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 InternVL3 5 4B HF?

4.7B parameters: 8.8 GB of weights at 16-bit, 2.7 GB at 4-bit.

Can InternVL3 5 4B HF run on an RTX 4090?

Yes. At 8-bit it holds 40K of context and writes about 131 tokens a second.

More models