Skip to content

InternVL3 78B VRAM requirements

How much VRAM InternVL3 78B needs
4-bit weights: 44.3 GBWith 8K context: 47.8 GBWith 32K context: 55.3 GB8-bit, 8K context: 81.1 GB
Parameters
78.4B
Longest context
32K
Released
Apr 2025
License
Qwen
Strongest at
Text and images

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

How much VRAM InternVL3 78B needs

Memory InternVL3 78B needs at each precision and context length
PrecisionWeightsWith 8KWith 32K
16-bit146 GB150 GB157 GB
8-bit77.6 GB81.1 GB88.6 GB
6-bit59.9 GB63.4 GB70.9 GB
5-bit51.9 GB55.4 GB62.9 GB
4-bit44.3 GB47.8 GB55.3 GB
3-bit35.7 GB39.2 GB46.7 GB
2-bit27.4 GB30.9 GB38.4 GB
1.6-bit15.1 GB18.6 GB26.1 GB

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

GPUs and Macs that run InternVL3 78B

The best way InternVL3 78B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: InternVL3 78B on RTX 4060 report
RTX 5060Too bigPDF: InternVL3 78B on RTX 5060 report
RTX 3070Too bigPDF: InternVL3 78B on RTX 3070 report
RTX 3080 10 GBToo bigPDF: InternVL3 78B on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: InternVL3 78B on RTX 3060 12 GB report
Arc B580Too bigPDF: InternVL3 78B on Arc B580 report
RTX 4070Too bigPDF: InternVL3 78B on RTX 4070 report
RTX 4070 SuperToo bigPDF: InternVL3 78B on RTX 4070 Super report
RTX 5070Too bigPDF: InternVL3 78B on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: InternVL3 78B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: InternVL3 78B on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: InternVL3 78B on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: InternVL3 78B on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: InternVL3 78B on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: InternVL3 78B on RTX 4080 Super report
RTX 5070 TiToo bigPDF: InternVL3 78B on RTX 5070 Ti report
RTX 5080Too bigPDF: InternVL3 78B on RTX 5080 report
Radeon RX 7900 XTXToo bigPDF: InternVL3 78B on Radeon RX 7900 XTX report
RTX 3090Too bigPDF: InternVL3 78B on RTX 3090 report
RTX 4090Too bigPDF: InternVL3 78B on RTX 4090 report
Radeon AI PRO R9700Too bigPDF: InternVL3 78B on Radeon AI PRO R9700 report
RTX 5090Too bigPDF: InternVL3 78B on RTX 5090 report
2× RTX 30903-bit24K17 tok/s2.7 sPDF: InternVL3 78B on 2× RTX 3090 report
2× RTX 40903-bit24K18 tok/s1.2 sPDF: InternVL3 78B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB3-bit28K24 tok/s2.5 sPDF: InternVL3 78B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada3-bit28K17 tok/s1.7 sPDF: InternVL3 78B on RTX 6000 Ada report
L40S3-bit28K15 tok/s2.9 sPDF: InternVL3 78B on L40S report
2× RTX 50905-bit20K22 tok/s0.9 sPDF: InternVL3 78B on 2× RTX 5090 report
4× RTX 30908-bit32K7.8 tok/s1.4 sPDF: InternVL3 78B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit32K15 tok/s1.3 sPDF: InternVL3 78B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: InternVL3 78B on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: InternVL3 78B on Mac mini M4 24 GB report
Mac mini M4 32 GBToo bigPDF: InternVL3 78B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: InternVL3 78B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GBToo bigPDF: InternVL3 78B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB3-bit28K5.7 tok/s53 sPDF: InternVL3 78B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit32K4.8 tok/s26 sPDF: InternVL3 78B on M4 Max 128 GB report
M5 Max 128 GB8-bit32K5.4 tok/s7.7 sPDF: InternVL3 78B on M5 Max 128 GB report
M3 Ultra 96 GB6-bit24K8.0 tok/s15 sPDF: InternVL3 78B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit32K6.2 tok/s15 sPDF: InternVL3 78B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit32K6.2 tok/s15 sPDF: InternVL3 78B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit32K9.0 tok/s4.6 sPDF: InternVL3 78B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB3-bit28K4.0 tok/s15 sPDF: InternVL3 78B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit32K1.9 tok/s15 sPDF: InternVL3 78B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit32K2.5 tok/s5.3 sPDF: InternVL3 78B on RTX Spark report
DGX Spark8-bit32K2.3 tok/s5.3 sPDF: InternVL3 78B on DGX Spark report
Dual DGX Spark8-bit32K2.3 tok/s3.0 sPDF: InternVL3 78B on Dual DGX Spark report
DGX Station GB3008-bit32K53 tok/s0.4 sPDF: InternVL3 78B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit32K15 tok/s0.7 sPDF: InternVL3 78B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit32K15 tok/s0.4 sPDF: InternVL3 78B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit32K15 tok/s0.2 sPDF: InternVL3 78B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM6-bit32K21 tok/s3.4 sPDF: InternVL3 78B on A100 80 GB SXM report
H100 SXM6-bit32K34 tok/s1.1 sPDF: InternVL3 78B on H100 SXM report
4× H200 NVL8-bit32K37 tok/s0.4 sPDF: InternVL3 78B on 4× H200 NVL report
8× H200 NVL8-bit32K37 tok/s0.2 sPDF: InternVL3 78B 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/InternVL3-78B-GGUF
vLLM
vllm serve OpenGVLab/InternVL3-78B

Questions

How much VRAM does InternVL3 78B need?

InternVL3 78B needs 44.3 GB for its weights at 4-bit, 47.8 GB with an 8K context and 55.3 GB with 32K. At 8-bit, 81.1 GB.

What is the minimum VRAM for InternVL3 78B?

About 39.2 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 2× RTX 3090.

How big is InternVL3 78B?

78.4B parameters: 146 GB of weights at 16-bit, 44.3 GB at 4-bit.

Can InternVL3 78B run on an RTX 4090?

Not on one card. The smallest machine here that runs it is the 2× RTX 3090.

More models