Skip to content

InternVL2 5 4B MPO VRAM requirements

How much VRAM InternVL2 5 4B MPO needs
4-bit weights: 2.1 GBWith 8K context: 3.2 GBWith 32K context: 4.0 GB8-bit, 8K context: 4.8 GB
Parameters
3.7B
Longest context
32K
Released
Dec 2024
License
MIT
Strongest at
Text and images

How much VRAM InternVL2 5 4B MPO needs

Memory InternVL2 5 4B MPO needs at each precision and context length
PrecisionWeightsWith 8KWith 32K
16-bit6.9 GB8.0 GB8.8 GB
8-bit3.7 GB4.8 GB5.6 GB
6-bit2.8 GB3.9 GB4.8 GB
5-bit2.5 GB3.5 GB4.4 GB
4-bit2.1 GB3.2 GB4.0 GB
3-bit1.7 GB2.8 GB3.6 GB
2-bit1.3 GB2.4 GB3.2 GB
1.6-bit730 MB1.8 GB2.6 GB

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

GPUs and Macs that run InternVL2 5 4B MPO

The best way InternVL2 5 4B MPO runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40608-bit32K50 tok/s0.5 sPDF: InternVL2 5 4B MPO on RTX 4060 report
RTX 50608-bit32K79 tok/s0.4 sPDF: InternVL2 5 4B MPO on RTX 5060 report
RTX 30708-bit32K79 tok/s0.3 sPDF: InternVL2 5 4B MPO on RTX 3070 report
RTX 3080 10 GB8-bit32K127 tok/s0.2 sPDF: InternVL2 5 4B MPO on RTX 3080 10 GB report
RTX 3060 12 GB8-bit32K65 tok/s0.5 sPDF: InternVL2 5 4B MPO on RTX 3060 12 GB report
Arc B5808-bit32K56 tok/s0.4 sPDF: InternVL2 5 4B MPO on Arc B580 report
RTX 40708-bit32K88 tok/s0.2 sPDF: InternVL2 5 4B MPO on RTX 4070 report
RTX 4070 Super8-bit32K88 tok/s0.2 sPDF: InternVL2 5 4B MPO on RTX 4070 Super report
RTX 50708-bit32K114 tok/s0.2 sPDF: InternVL2 5 4B MPO on RTX 5070 report
RTX 4060 Ti 16 GB8-bit32K53 tok/s0.3 sPDF: InternVL2 5 4B MPO on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit32K79 tok/s0.3 sPDF: InternVL2 5 4B MPO on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit32K66 tok/s0.3 sPDF: InternVL2 5 4B MPO on Arc A770 16 GB report
Radeon RX 9070 XT8-bit32K89 tok/s0.2 sPDF: InternVL2 5 4B MPO on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit32K114 tok/s0.2 sPDF: InternVL2 5 4B MPO on RTX 4070 Ti Super report
RTX 4080 Super8-bit32K124 tok/s0.1 sPDF: InternVL2 5 4B MPO on RTX 4080 Super report
RTX 5070 Ti8-bit32K146 tok/s0.2 sPDF: InternVL2 5 4B MPO on RTX 5070 Ti report
RTX 50808-bit32K155 tok/s0.1 sPDF: InternVL2 5 4B MPO on RTX 5080 report
Radeon RX 7900 XTX8-bit32K124 tok/s0.3 sPDF: InternVL2 5 4B MPO on Radeon RX 7900 XTX report
RTX 30908-bit32K152 tok/s0.2 sPDF: InternVL2 5 4B MPO on RTX 3090 report
RTX 40908-bit32K161 tok/s0.1 sPDF: InternVL2 5 4B MPO on RTX 4090 report
Radeon AI PRO R97008-bit32K89 tok/s0.2 sPDF: InternVL2 5 4B MPO on Radeon AI PRO R9700 report
RTX 50908-bit32K252 tok/s0.1 sPDF: InternVL2 5 4B MPO on RTX 5090 report
2× RTX 30908-bit32K152 tok/s0.1 sPDF: InternVL2 5 4B MPO on 2× RTX 3090 report
2× RTX 40908-bit32K161 tok/s0.1 sPDF: InternVL2 5 4B MPO on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit32K203 tok/s0.1 sPDF: InternVL2 5 4B MPO on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit32K155 tok/s0.1 sPDF: InternVL2 5 4B MPO on RTX 6000 Ada report
L40S8-bit32K142 tok/s0.1 sPDF: InternVL2 5 4B MPO on L40S report
2× RTX 50908-bit32K252 tok/s0.1 sPDF: InternVL2 5 4B MPO on 2× RTX 5090 report
4× RTX 30908-bit32K152 tok/s0.1 sPDF: InternVL2 5 4B MPO on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit32K252 tok/s0.1 sPDF: InternVL2 5 4B MPO on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit32K25 tok/s4.4 sPDF: InternVL2 5 4B MPO on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit32K25 tok/s4.4 sPDF: InternVL2 5 4B MPO on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit32K25 tok/s4.4 sPDF: InternVL2 5 4B MPO on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit32K53 tok/s2.2 sPDF: InternVL2 5 4B MPO on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit32K53 tok/s2.2 sPDF: InternVL2 5 4B MPO on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit32K53 tok/s2.2 sPDF: InternVL2 5 4B MPO on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit32K85 tok/s1.1 sPDF: InternVL2 5 4B MPO on M4 Max 128 GB report
M5 Max 128 GB8-bit32K93 tok/s0.3 sPDF: InternVL2 5 4B MPO on M5 Max 128 GB report
M3 Ultra 96 GB8-bit32K103 tok/s0.6 sPDF: InternVL2 5 4B MPO on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit32K103 tok/s0.6 sPDF: InternVL2 5 4B MPO on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit32K103 tok/s0.6 sPDF: InternVL2 5 4B MPO on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit32K133 tok/s0.2 sPDF: InternVL2 5 4B MPO on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit32K39 tok/s0.6 sPDF: InternVL2 5 4B MPO on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit32K39 tok/s0.6 sPDF: InternVL2 5 4B MPO on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit32K55 tok/s0.2 sPDF: InternVL2 5 4B MPO on RTX Spark report
DGX Spark8-bit32K50 tok/s0.2 sPDF: InternVL2 5 4B MPO on DGX Spark report
Dual DGX Spark8-bit32K50 tok/s0.1 sPDF: InternVL2 5 4B MPO on Dual DGX Spark report
DGX Station GB3008-bit32K553 tok/s0.1 sPDF: InternVL2 5 4B MPO on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit32K252 tok/s0.1 sPDF: InternVL2 5 4B MPO on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit32K252 tok/s0.1 sPDF: InternVL2 5 4B MPO on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit32K252 tok/s0.1 sPDF: InternVL2 5 4B MPO on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit32K277 tok/s0.1 sPDF: InternVL2 5 4B MPO on A100 80 GB SXM report
H100 SXM8-bit32K381 tok/s0.1 sPDF: InternVL2 5 4B MPO on H100 SXM report
4× H200 NVL8-bit32K464 tok/s0.1 sPDF: InternVL2 5 4B MPO on 4× H200 NVL report
8× H200 NVL8-bit32K464 tok/s0.1 sPDF: InternVL2 5 4B MPO 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 mradermacher/InternVL2_5-4B-MPO-GGUF
vLLM
vllm serve OpenGVLab/InternVL2_5-4B-MPO

Questions

How much VRAM does InternVL2 5 4B MPO need?

InternVL2 5 4B MPO needs 2.1 GB for its weights at 4-bit, 3.2 GB with an 8K context and 4.0 GB with 32K. At 8-bit, 4.8 GB.

What is the minimum VRAM for InternVL2 5 4B MPO?

About 2.8 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 InternVL2 5 4B MPO?

3.7B parameters: 6.9 GB of weights at 16-bit, 2.1 GB at 4-bit.

Can InternVL2 5 4B MPO run on an RTX 4090?

Yes. At 8-bit it holds 32K of context and writes about 161 tokens a second.

More models