Skip to content

Idefics3 8B Llama3 VRAM requirements

How much VRAM Idefics3 8B Llama3 needs
4-bit weights: 4.8 GBWith 8K context: 6.6 GBWith 32K context: 9.6 GB8-bit, 8K context: 10.2 GB
Parameters
8.5B
Longest context
128K
Released
Aug 2024
License
Apache 2.0
Strongest at
Text and images

How much VRAM Idefics3 8B Llama3 needs

Memory Idefics3 8B Llama3 needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit15.8 GB17.6 GB20.6 GB32.6 GB
8-bit8.4 GB10.2 GB13.2 GB25.2 GB
6-bit6.5 GB8.3 GB11.3 GB23.3 GB
5-bit5.6 GB7.4 GB10.4 GB22.4 GB
4-bit4.8 GB6.6 GB9.6 GB21.6 GB
3-bit3.9 GB5.7 GB8.7 GB20.7 GB
2-bit3.0 GB4.8 GB7.8 GB19.8 GB
1.6-bit1.6 GB3.4 GB6.4 GB18.4 GB

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

GPUs and Macs that run Idefics3 8B Llama3

The best way Idefics3 8B Llama3 runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40603-bit20K47 tok/s1.1 sPDF: Idefics3 8B Llama3 on RTX 4060 report
RTX 50603-bit20K75 tok/s0.9 sPDF: Idefics3 8B Llama3 on RTX 5060 report
RTX 30703-bit20K75 tok/s0.8 sPDF: Idefics3 8B Llama3 on RTX 3070 report
RTX 3080 10 GB6-bit16K76 tok/s0.6 sPDF: Idefics3 8B Llama3 on RTX 3080 10 GB report
RTX 3060 12 GB8-bit16K29 tok/s1.3 sPDF: Idefics3 8B Llama3 on RTX 3060 12 GB report
Arc B5808-bit16K26 tok/s0.9 sPDF: Idefics3 8B Llama3 on Arc B580 report
RTX 40708-bit16K40 tok/s0.6 sPDF: Idefics3 8B Llama3 on RTX 4070 report
RTX 4070 Super8-bit16K40 tok/s0.5 sPDF: Idefics3 8B Llama3 on RTX 4070 Super report
RTX 50708-bit16K53 tok/s0.5 sPDF: Idefics3 8B Llama3 on RTX 5070 report
RTX 4060 Ti 16 GB8-bit48K23 tok/s0.8 sPDF: Idefics3 8B Llama3 on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit48K36 tok/s0.7 sPDF: Idefics3 8B Llama3 on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit48K31 tok/s0.8 sPDF: Idefics3 8B Llama3 on Arc A770 16 GB report
Radeon RX 9070 XT8-bit48K42 tok/s0.5 sPDF: Idefics3 8B Llama3 on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit48K53 tok/s0.4 sPDF: Idefics3 8B Llama3 on RTX 4070 Ti Super report
RTX 4080 Super8-bit48K58 tok/s0.3 sPDF: Idefics3 8B Llama3 on RTX 4080 Super report
RTX 5070 Ti8-bit48K69 tok/s0.4 sPDF: Idefics3 8B Llama3 on RTX 5070 Ti report
RTX 50808-bit48K74 tok/s0.3 sPDF: Idefics3 8B Llama3 on RTX 5080 report
Radeon RX 7900 XTX8-bit104K62 tok/s0.7 sPDF: Idefics3 8B Llama3 on Radeon RX 7900 XTX report
RTX 30908-bit104K72 tok/s0.5 sPDF: Idefics3 8B Llama3 on RTX 3090 report
RTX 40908-bit104K77 tok/s0.2 sPDF: Idefics3 8B Llama3 on RTX 4090 report
Radeon AI PRO R97008-bit128K42 tok/s0.5 sPDF: Idefics3 8B Llama3 on Radeon AI PRO R9700 report
RTX 50908-bit128K130 tok/s0.2 sPDF: Idefics3 8B Llama3 on RTX 5090 report
2× RTX 30908-bit128K72 tok/s0.3 sPDF: Idefics3 8B Llama3 on 2× RTX 3090 report
2× RTX 40908-bit128K77 tok/s0.1 sPDF: Idefics3 8B Llama3 on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit128K101 tok/s0.2 sPDF: Idefics3 8B Llama3 on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit128K74 tok/s0.2 sPDF: Idefics3 8B Llama3 on RTX 6000 Ada report
L40S8-bit128K67 tok/s0.3 sPDF: Idefics3 8B Llama3 on L40S report
2× RTX 50908-bit128K130 tok/s0.1 sPDF: Idefics3 8B Llama3 on 2× RTX 5090 report
4× RTX 30908-bit128K72 tok/s0.1 sPDF: Idefics3 8B Llama3 on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit128K130 tok/s0.1 sPDF: Idefics3 8B Llama3 on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB6-bit24K15 tok/s11 sPDF: Idefics3 8B Llama3 on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit48K11 tok/s11 sPDF: Idefics3 8B Llama3 on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit88K11 tok/s11 sPDF: Idefics3 8B Llama3 on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit48K25 tok/s5.3 sPDF: Idefics3 8B Llama3 on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit128K25 tok/s5.3 sPDF: Idefics3 8B Llama3 on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit128K25 tok/s5.3 sPDF: Idefics3 8B Llama3 on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit128K43 tok/s2.6 sPDF: Idefics3 8B Llama3 on M4 Max 128 GB report
M5 Max 128 GB8-bit128K48 tok/s0.8 sPDF: Idefics3 8B Llama3 on M5 Max 128 GB report
M3 Ultra 96 GB8-bit128K54 tok/s1.5 sPDF: Idefics3 8B Llama3 on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit128K54 tok/s1.5 sPDF: Idefics3 8B Llama3 on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit128K54 tok/s1.5 sPDF: Idefics3 8B Llama3 on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit128K75 tok/s0.5 sPDF: Idefics3 8B Llama3 on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit128K18 tok/s1.5 sPDF: Idefics3 8B Llama3 on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit128K18 tok/s1.5 sPDF: Idefics3 8B Llama3 on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit128K24 tok/s0.5 sPDF: Idefics3 8B Llama3 on RTX Spark report
DGX Spark8-bit128K22 tok/s0.5 sPDF: Idefics3 8B Llama3 on DGX Spark report
Dual DGX Spark8-bit128K22 tok/s0.3 sPDF: Idefics3 8B Llama3 on Dual DGX Spark report
DGX Station GB3008-bit128K376 tok/s0.1 sPDF: Idefics3 8B Llama3 on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit128K130 tok/s0.1 sPDF: Idefics3 8B Llama3 on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit128K130 tok/s0.1 sPDF: Idefics3 8B Llama3 on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit128K130 tok/s0.1 sPDF: Idefics3 8B Llama3 on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit128K146 tok/s0.3 sPDF: Idefics3 8B Llama3 on A100 80 GB SXM report
H100 SXM8-bit128K219 tok/s0.1 sPDF: Idefics3 8B Llama3 on H100 SXM report
4× H200 NVL8-bit128K288 tok/s0.1 sPDF: Idefics3 8B Llama3 on 4× H200 NVL report
8× H200 NVL8-bit128K288 tok/s0.1 sPDF: Idefics3 8B Llama3 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 jesse9527/Idefics3-8B-Llama3-Q4_K_M-GGUF
vLLM
vllm serve HuggingFaceM4/Idefics3-8B-Llama3

Questions

How much VRAM does Idefics3 8B Llama3 need?

Idefics3 8B Llama3 needs 4.8 GB for its weights at 4-bit, 6.6 GB with an 8K context and 9.6 GB with 32K. At 8-bit, 10.2 GB.

What is the minimum VRAM for Idefics3 8B Llama3?

About 5.7 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 Idefics3 8B Llama3?

8.5B parameters: 15.8 GB of weights at 16-bit, 4.8 GB at 4-bit.

Can Idefics3 8B Llama3 run on an RTX 4090?

Yes. At 8-bit it holds 104K of context and writes about 77 tokens a second.

More models