Skip to content

AIN VRAM requirements

How much VRAM AIN needs
4-bit weights: 4.7 GBWith 8K context: 6.0 GBWith 32K context: 7.3 GB8-bit, 8K context: 9.5 GB
Parameters
8.3B
Longest context
32K
Released
Jan 2025
License
MIT
Strongest at
Text and images

How much VRAM AIN needs

Memory AIN needs at each precision and context length
PrecisionWeightsWith 8KWith 32K
16-bit15.4 GB16.7 GB18.0 GB
8-bit8.2 GB9.5 GB10.8 GB
6-bit6.3 GB7.6 GB8.9 GB
5-bit5.5 GB6.8 GB8.1 GB
4-bit4.7 GB6.0 GB7.3 GB
3-bit3.8 GB5.1 GB6.4 GB
2-bit2.9 GB4.2 GB5.5 GB
1.6-bit1.6 GB2.9 GB4.2 GB

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

GPUs and Macs that run AIN

The best way AIN runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40605-bit20K34 tok/s1.1 sPDF: AIN on RTX 4060 report
RTX 50605-bit20K54 tok/s0.8 sPDF: AIN on RTX 5060 report
RTX 30705-bit20K54 tok/s0.8 sPDF: AIN on RTX 3070 report
RTX 3080 10 GB6-bit32K78 tok/s0.5 sPDF: AIN on RTX 3080 10 GB report
RTX 3060 12 GB8-bit32K30 tok/s1.3 sPDF: AIN on RTX 3060 12 GB report
Arc B5808-bit32K26 tok/s0.9 sPDF: AIN on Arc B580 report
RTX 40708-bit32K41 tok/s0.5 sPDF: AIN on RTX 4070 report
RTX 4070 Super8-bit32K41 tok/s0.4 sPDF: AIN on RTX 4070 Super report
RTX 50708-bit32K55 tok/s0.5 sPDF: AIN on RTX 5070 report
RTX 4060 Ti 16 GB8-bit32K24 tok/s0.7 sPDF: AIN on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit32K37 tok/s0.7 sPDF: AIN on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit32K32 tok/s0.8 sPDF: AIN on Arc A770 16 GB report
Radeon RX 9070 XT8-bit32K44 tok/s0.4 sPDF: AIN on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit32K55 tok/s0.4 sPDF: AIN on RTX 4070 Ti Super report
RTX 4080 Super8-bit32K59 tok/s0.3 sPDF: AIN on RTX 4080 Super report
RTX 5070 Ti8-bit32K72 tok/s0.4 sPDF: AIN on RTX 5070 Ti report
RTX 50808-bit32K76 tok/s0.3 sPDF: AIN on RTX 5080 report
Radeon RX 7900 XTX8-bit32K64 tok/s0.7 sPDF: AIN on Radeon RX 7900 XTX report
RTX 30908-bit32K75 tok/s0.4 sPDF: AIN on RTX 3090 report
RTX 40908-bit32K80 tok/s0.2 sPDF: AIN on RTX 4090 report
Radeon AI PRO R97008-bit32K44 tok/s0.5 sPDF: AIN on Radeon AI PRO R9700 report
RTX 50908-bit32K135 tok/s0.2 sPDF: AIN on RTX 5090 report
2× RTX 30908-bit32K75 tok/s0.3 sPDF: AIN on 2× RTX 3090 report
2× RTX 40908-bit32K80 tok/s0.1 sPDF: AIN on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit32K104 tok/s0.2 sPDF: AIN on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit32K76 tok/s0.2 sPDF: AIN on RTX 6000 Ada report
L40S8-bit32K69 tok/s0.3 sPDF: AIN on L40S report
2× RTX 50908-bit32K135 tok/s0.1 sPDF: AIN on 2× RTX 5090 report
4× RTX 30908-bit32K75 tok/s0.1 sPDF: AIN on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit32K135 tok/s0.1 sPDF: AIN on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit20K12 tok/s10 sPDF: AIN on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit32K12 tok/s10 sPDF: AIN on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit32K12 tok/s10 sPDF: AIN on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit32K26 tok/s5.1 sPDF: AIN on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit32K26 tok/s5.1 sPDF: AIN on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit32K26 tok/s5.1 sPDF: AIN on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit32K45 tok/s2.5 sPDF: AIN on M4 Max 128 GB report
M5 Max 128 GB8-bit32K50 tok/s0.7 sPDF: AIN on M5 Max 128 GB report
M3 Ultra 96 GB8-bit32K56 tok/s1.4 sPDF: AIN on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit32K56 tok/s1.4 sPDF: AIN on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit32K56 tok/s1.4 sPDF: AIN on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit32K78 tok/s0.4 sPDF: AIN on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit32K18 tok/s1.5 sPDF: AIN on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit32K18 tok/s1.5 sPDF: AIN on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit32K25 tok/s0.5 sPDF: AIN on RTX Spark report
DGX Spark8-bit32K23 tok/s0.5 sPDF: AIN on DGX Spark report
Dual DGX Spark8-bit32K23 tok/s0.3 sPDF: AIN on Dual DGX Spark report
DGX Station GB3008-bit32K401 tok/s0.1 sPDF: AIN on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit32K135 tok/s0.1 sPDF: AIN on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit32K135 tok/s0.1 sPDF: AIN on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit32K135 tok/s0.1 sPDF: AIN on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit32K151 tok/s0.3 sPDF: AIN on A100 80 GB SXM report
H100 SXM8-bit32K230 tok/s0.1 sPDF: AIN on H100 SXM report
4× H200 NVL8-bit32K304 tok/s0.1 sPDF: AIN on 4× H200 NVL report
8× H200 NVL8-bit32K304 tok/s0.1 sPDF: AIN 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 zachq2122/MBZUAI-AIN-7B
vLLM
vllm serve MBZUAI/AIN

Questions

How much VRAM does AIN need?

AIN needs 4.7 GB for its weights at 4-bit, 6.0 GB with an 8K context and 7.3 GB with 32K. At 8-bit, 9.5 GB.

What is the minimum VRAM for AIN?

About 5.1 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 AIN?

8.3B parameters: 15.4 GB of weights at 16-bit, 4.7 GB at 4-bit.

Can AIN run on an RTX 4090?

Yes. At 8-bit it holds 32K of context and writes about 80 tokens a second.

More models