AIN VRAM requirements
| 4-bit weights: 4.7 GB | With 8K context: 6.0 GB | With 32K context: 7.3 GB | 8-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
| Precision | Weights | With 8K | With 32K |
|---|---|---|---|
| 16-bit | 15.4 GB | 16.7 GB | 18.0 GB |
| 8-bit | 8.2 GB | 9.5 GB | 10.8 GB |
| 6-bit | 6.3 GB | 7.6 GB | 8.9 GB |
| 5-bit | 5.5 GB | 6.8 GB | 8.1 GB |
| 4-bit | 4.7 GB | 6.0 GB | 7.3 GB |
| 3-bit | 3.8 GB | 5.1 GB | 6.4 GB |
| 2-bit | 2.9 GB | 4.2 GB | 5.5 GB |
| 1.6-bit | 1.6 GB | 2.9 GB | 4.2 GB |
One conversation, 16-bit context cache, runtime overhead included.
GPUs and Macs that run AIN
| Machine | Best fit | Context | Speed | First reply | Report |
|---|---|---|---|---|---|
| RTX 4060 | 5-bit | 20K | 34 tok/s | 1.1 s | PDF: AIN on RTX 4060 report |
| RTX 5060 | 5-bit | 20K | 54 tok/s | 0.8 s | PDF: AIN on RTX 5060 report |
| RTX 3070 | 5-bit | 20K | 54 tok/s | 0.8 s | PDF: AIN on RTX 3070 report |
| RTX 3080 10 GB | 6-bit | 32K | 78 tok/s | 0.5 s | PDF: AIN on RTX 3080 10 GB report |
| RTX 3060 12 GB | 8-bit | 32K | 30 tok/s | 1.3 s | PDF: AIN on RTX 3060 12 GB report |
| Arc B580 | 8-bit | 32K | 26 tok/s | 0.9 s | PDF: AIN on Arc B580 report |
| RTX 4070 | 8-bit | 32K | 41 tok/s | 0.5 s | PDF: AIN on RTX 4070 report |
| RTX 4070 Super | 8-bit | 32K | 41 tok/s | 0.4 s | PDF: AIN on RTX 4070 Super report |
| RTX 5070 | 8-bit | 32K | 55 tok/s | 0.5 s | PDF: AIN on RTX 5070 report |
| RTX 4060 Ti 16 GB | 8-bit | 32K | 24 tok/s | 0.7 s | PDF: AIN on RTX 4060 Ti 16 GB report |
| RTX 5060 Ti 16 GB | 8-bit | 32K | 37 tok/s | 0.7 s | PDF: AIN on RTX 5060 Ti 16 GB report |
| Arc A770 16 GB | 8-bit | 32K | 32 tok/s | 0.8 s | PDF: AIN on Arc A770 16 GB report |
| Radeon RX 9070 XT | 8-bit | 32K | 44 tok/s | 0.4 s | PDF: AIN on Radeon RX 9070 XT report |
| RTX 4070 Ti Super | 8-bit | 32K | 55 tok/s | 0.4 s | PDF: AIN on RTX 4070 Ti Super report |
| RTX 4080 Super | 8-bit | 32K | 59 tok/s | 0.3 s | PDF: AIN on RTX 4080 Super report |
| RTX 5070 Ti | 8-bit | 32K | 72 tok/s | 0.4 s | PDF: AIN on RTX 5070 Ti report |
| RTX 5080 | 8-bit | 32K | 76 tok/s | 0.3 s | PDF: AIN on RTX 5080 report |
| Radeon RX 7900 XTX | 8-bit | 32K | 64 tok/s | 0.7 s | PDF: AIN on Radeon RX 7900 XTX report |
| RTX 3090 | 8-bit | 32K | 75 tok/s | 0.4 s | PDF: AIN on RTX 3090 report |
| RTX 4090 | 8-bit | 32K | 80 tok/s | 0.2 s | PDF: AIN on RTX 4090 report |
| Radeon AI PRO R9700 | 8-bit | 32K | 44 tok/s | 0.5 s | PDF: AIN on Radeon AI PRO R9700 report |
| RTX 5090 | 8-bit | 32K | 135 tok/s | 0.2 s | PDF: AIN on RTX 5090 report |
| 2× RTX 3090 | 8-bit | 32K | 75 tok/s | 0.3 s | PDF: AIN on 2× RTX 3090 report |
| 2× RTX 4090 | 8-bit | 32K | 80 tok/s | 0.1 s | PDF: AIN on 2× RTX 4090 report |
| RTX PRO 5000 Blackwell 48 GB | 8-bit | 32K | 104 tok/s | 0.2 s | PDF: AIN on RTX PRO 5000 Blackwell 48 GB report |
| RTX 6000 Ada | 8-bit | 32K | 76 tok/s | 0.2 s | PDF: AIN on RTX 6000 Ada report |
| L40S | 8-bit | 32K | 69 tok/s | 0.3 s | PDF: AIN on L40S report |
| 2× RTX 5090 | 8-bit | 32K | 135 tok/s | 0.1 s | PDF: AIN on 2× RTX 5090 report |
| 4× RTX 3090 | 8-bit | 32K | 75 tok/s | 0.1 s | PDF: AIN on 4× RTX 3090 report |
| RTX PRO 6000 Blackwell | 8-bit | 32K | 135 tok/s | 0.1 s | PDF: AIN on RTX PRO 6000 Blackwell report |
| Mac mini M4 16 GB | 8-bit | 20K | 12 tok/s | 10 s | PDF: AIN on Mac mini M4 16 GB report |
| Mac mini M4 24 GB | 8-bit | 32K | 12 tok/s | 10 s | PDF: AIN on Mac mini M4 24 GB report |
| Mac mini M4 32 GB | 8-bit | 32K | 12 tok/s | 10 s | PDF: AIN on Mac mini M4 32 GB report |
| Mac mini M4 Pro 24 GB | 8-bit | 32K | 26 tok/s | 5.1 s | PDF: AIN on Mac mini M4 Pro 24 GB report |
| Mac mini M4 Pro 48 GB | 8-bit | 32K | 26 tok/s | 5.1 s | PDF: AIN on Mac mini M4 Pro 48 GB report |
| Mac mini M4 Pro 64 GB | 8-bit | 32K | 26 tok/s | 5.1 s | PDF: AIN on Mac mini M4 Pro 64 GB report |
| M4 Max 128 GB | 8-bit | 32K | 45 tok/s | 2.5 s | PDF: AIN on M4 Max 128 GB report |
| M5 Max 128 GB | 8-bit | 32K | 50 tok/s | 0.7 s | PDF: AIN on M5 Max 128 GB report |
| M3 Ultra 96 GB | 8-bit | 32K | 56 tok/s | 1.4 s | PDF: AIN on M3 Ultra 96 GB report |
| M3 Ultra 256 GB | 8-bit | 32K | 56 tok/s | 1.4 s | PDF: AIN on M3 Ultra 256 GB report |
| M3 Ultra 512 GB | 8-bit | 32K | 56 tok/s | 1.4 s | PDF: AIN on M3 Ultra 512 GB report |
| M5 Ultra 512 GB | 8-bit | 32K | 78 tok/s | 0.4 s | PDF: AIN on M5 Ultra 512 GB report |
| Ryzen AI Max+ 395 64 GB | 8-bit | 32K | 18 tok/s | 1.5 s | PDF: AIN on Ryzen AI Max+ 395 64 GB report |
| Ryzen AI Max+ 395 128 GB | 8-bit | 32K | 18 tok/s | 1.5 s | PDF: AIN on Ryzen AI Max+ 395 128 GB report |
| RTX Spark | 8-bit | 32K | 25 tok/s | 0.5 s | PDF: AIN on RTX Spark report |
| DGX Spark | 8-bit | 32K | 23 tok/s | 0.5 s | PDF: AIN on DGX Spark report |
| Dual DGX Spark | 8-bit | 32K | 23 tok/s | 0.3 s | PDF: AIN on Dual DGX Spark report |
| DGX Station GB300 | 8-bit | 32K | 401 tok/s | 0.1 s | PDF: AIN on DGX Station GB300 report |
| 2× RTX PRO 6000 Blackwell | 8-bit | 32K | 135 tok/s | 0.1 s | PDF: AIN on 2× RTX PRO 6000 Blackwell report |
| 4× RTX PRO 6000 Blackwell | 8-bit | 32K | 135 tok/s | 0.1 s | PDF: AIN on 4× RTX PRO 6000 Blackwell report |
| 8× RTX PRO 6000 Blackwell | 8-bit | 32K | 135 tok/s | 0.1 s | PDF: AIN on 8× RTX PRO 6000 Blackwell report |
| A100 80 GB SXM | 8-bit | 32K | 151 tok/s | 0.3 s | PDF: AIN on A100 80 GB SXM report |
| H100 SXM | 8-bit | 32K | 230 tok/s | 0.1 s | PDF: AIN on H100 SXM report |
| 4× H200 NVL | 8-bit | 32K | 304 tok/s | 0.1 s | PDF: AIN on 4× H200 NVL report |
| 8× H200 NVL | 8-bit | 32K | 304 tok/s | 0.1 s | PDF: 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.