Skip to content

InternVL3 5 8B VRAM requirements

How much VRAM InternVL3 5 8B needs
4-bit weights: 4.8 GBWith 8K context: 6.8 GBWith 32K context: 10.2 GB8-bit, 8K context: 10.4 GB
Parameters
8.5B
Longest context
40K
Released
Aug 2025
License
Apache 2.0
Strongest at
Text and images

How much VRAM InternVL3 5 8B needs

Memory InternVL3 5 8B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 40K
16-bit15.9 GB17.9 GB21.3 GB22.4 GB
8-bit8.4 GB10.4 GB13.8 GB14.9 GB
6-bit6.5 GB8.5 GB11.9 GB13.0 GB
5-bit5.6 GB7.6 GB11.0 GB12.1 GB
4-bit4.8 GB6.8 GB10.2 GB11.3 GB
3-bit3.9 GB5.9 GB9.2 GB10.4 GB
2-bit3.0 GB5.0 GB8.3 GB9.5 GB
1.6-bit1.6 GB3.6 GB7.0 GB8.1 GB

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

GPUs and Macs that run InternVL3 5 8B

The best way InternVL3 5 8B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40603-bit20K47 tok/s1.1 sPDF: InternVL3 5 8B on RTX 4060 report
RTX 50603-bit20K75 tok/s0.9 sPDF: InternVL3 5 8B on RTX 5060 report
RTX 30703-bit20K75 tok/s0.8 sPDF: InternVL3 5 8B on RTX 3070 report
RTX 3080 10 GB5-bit20K86 tok/s0.5 sPDF: InternVL3 5 8B on RTX 3080 10 GB report
RTX 3060 12 GB6-bit28K37 tok/s1.3 sPDF: InternVL3 5 8B on RTX 3060 12 GB report
Arc B5806-bit28K33 tok/s0.9 sPDF: InternVL3 5 8B on Arc B580 report
RTX 40706-bit28K51 tok/s0.6 sPDF: InternVL3 5 8B on RTX 4070 report
RTX 4070 Super6-bit28K51 tok/s0.5 sPDF: InternVL3 5 8B on RTX 4070 Super report
RTX 50706-bit28K67 tok/s0.5 sPDF: InternVL3 5 8B on RTX 5070 report
RTX 4060 Ti 16 GB8-bit40K23 tok/s0.7 sPDF: InternVL3 5 8B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit40K36 tok/s0.7 sPDF: InternVL3 5 8B on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit40K31 tok/s0.8 sPDF: InternVL3 5 8B on Arc A770 16 GB report
Radeon RX 9070 XT8-bit40K42 tok/s0.5 sPDF: InternVL3 5 8B on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit40K53 tok/s0.4 sPDF: InternVL3 5 8B on RTX 4070 Ti Super report
RTX 4080 Super8-bit40K58 tok/s0.3 sPDF: InternVL3 5 8B on RTX 4080 Super report
RTX 5070 Ti8-bit40K69 tok/s0.4 sPDF: InternVL3 5 8B on RTX 5070 Ti report
RTX 50808-bit40K74 tok/s0.3 sPDF: InternVL3 5 8B on RTX 5080 report
Radeon RX 7900 XTX8-bit40K61 tok/s0.7 sPDF: InternVL3 5 8B on Radeon RX 7900 XTX report
RTX 30908-bit40K72 tok/s0.5 sPDF: InternVL3 5 8B on RTX 3090 report
RTX 40908-bit40K77 tok/s0.2 sPDF: InternVL3 5 8B on RTX 4090 report
Radeon AI PRO R97008-bit40K42 tok/s0.5 sPDF: InternVL3 5 8B on Radeon AI PRO R9700 report
RTX 50908-bit40K129 tok/s0.2 sPDF: InternVL3 5 8B on RTX 5090 report
2× RTX 30908-bit40K72 tok/s0.3 sPDF: InternVL3 5 8B on 2× RTX 3090 report
2× RTX 40908-bit40K77 tok/s0.1 sPDF: InternVL3 5 8B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit40K100 tok/s0.2 sPDF: InternVL3 5 8B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit40K74 tok/s0.2 sPDF: InternVL3 5 8B on RTX 6000 Ada report
L40S8-bit40K67 tok/s0.3 sPDF: InternVL3 5 8B on L40S report
2× RTX 50908-bit40K129 tok/s0.1 sPDF: InternVL3 5 8B on 2× RTX 5090 report
4× RTX 30908-bit40K72 tok/s0.1 sPDF: InternVL3 5 8B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit40K129 tok/s0.1 sPDF: InternVL3 5 8B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB6-bit20K15 tok/s10 sPDF: InternVL3 5 8B on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit40K11 tok/s10 sPDF: InternVL3 5 8B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit40K11 tok/s10 sPDF: InternVL3 5 8B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit40K25 tok/s5.2 sPDF: InternVL3 5 8B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit40K25 tok/s5.2 sPDF: InternVL3 5 8B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit40K25 tok/s5.2 sPDF: InternVL3 5 8B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit40K43 tok/s2.6 sPDF: InternVL3 5 8B on M4 Max 128 GB report
M5 Max 128 GB8-bit40K47 tok/s0.8 sPDF: InternVL3 5 8B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit40K54 tok/s1.5 sPDF: InternVL3 5 8B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit40K54 tok/s1.5 sPDF: InternVL3 5 8B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit40K54 tok/s1.5 sPDF: InternVL3 5 8B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit40K73 tok/s0.5 sPDF: InternVL3 5 8B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit40K18 tok/s1.5 sPDF: InternVL3 5 8B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit40K18 tok/s1.5 sPDF: InternVL3 5 8B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit40K24 tok/s0.5 sPDF: InternVL3 5 8B on RTX Spark report
DGX Spark8-bit40K22 tok/s0.5 sPDF: InternVL3 5 8B on DGX Spark report
Dual DGX Spark8-bit40K22 tok/s0.3 sPDF: InternVL3 5 8B on Dual DGX Spark report
DGX Station GB3008-bit40K361 tok/s0.1 sPDF: InternVL3 5 8B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit40K129 tok/s0.1 sPDF: InternVL3 5 8B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit40K129 tok/s0.1 sPDF: InternVL3 5 8B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit40K129 tok/s0.1 sPDF: InternVL3 5 8B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit40K144 tok/s0.3 sPDF: InternVL3 5 8B on A100 80 GB SXM report
H100 SXM8-bit40K214 tok/s0.1 sPDF: InternVL3 5 8B on H100 SXM report
4× H200 NVL8-bit40K279 tok/s0.1 sPDF: InternVL3 5 8B on 4× H200 NVL report
8× H200 NVL8-bit40K279 tok/s0.1 sPDF: InternVL3 5 8B 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 bartowski/OpenGVLab_InternVL3_5-8B-GGUF
vLLM
vllm serve OpenGVLab/InternVL3_5-8B

Questions

How much VRAM does InternVL3 5 8B need?

InternVL3 5 8B needs 4.8 GB for its weights at 4-bit, 6.8 GB with an 8K context and 10.2 GB with 32K. At 8-bit, 10.4 GB.

What is the minimum VRAM for InternVL3 5 8B?

About 5.9 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 8B?

8.5B parameters: 15.9 GB of weights at 16-bit, 4.8 GB at 4-bit.

Can InternVL3 5 8B run on an RTX 4090?

Yes. At 8-bit it holds 40K of context and writes about 77 tokens a second.

More models