# Hardware Prerequisites

The table below outlines the minimum and recommended memory requirements for running each Trinity model at different precision levels. In general, **4-bit and 8-bit quantization are ideal for most use cases**, offering a strong balance between performance and memory efficiency. These lower precisions make it easier to deploy on a wide range of hardware, including CPUs. 

| **Model**              | **Precision** | **Minimum RAM** | **Recommended RAM** |
| ---------------------- | ------------- | --------------- | ------------------- |
| **Trinity-Nano-6B**    | 4-bit         | 4 GB            | 6 GB                |
| **Trinity-Nano-6B**    | 8-bit         | 8 GB            | 8 GB                |
| **Trinity-Nano-6B**    | bf16          | 14 GB           | 14 GB               |
| **Trinity-Mini-26B**   | 4-bit         | 16 GB           | 24 GB               |
| **Trinity-Mini-26B**   | 8-bit         | 32 GB           | 32 GB               |
| **Trinity-Mini-26B**   | bf16          | 64 GB           | 64 GB               |
| **Trinity-Large-400B** | 4-bit         | 224GB           | 336GB               |
| **Trinity-Large-400B** | 8-bit         | 448GB           | 448GB               |
| **Trinity-Large-400B** | bf16          | 896GB           | 896GB               |

## CPU vs. GPU Considerations

GPUs are well suited for AI workloads that require high parallel processing power, such as training large models or handling high-throughput inference. They remain the preferred choice for tasks that demand maximum performance and low latency at scale.

CPUs, however, are becoming increasingly capable for AI inference due to improvements in hardware acceleration and model optimization. While not as fast as GPUs in most cases, CPUs offer lower power consumption, broader availability, and cost efficiency. For many real-time and interactive applications, optimized models running on CPUs can deliver adequate performance without the need for specialized hardware.

This makes CPUs a practical option for local deployments, edge devices, and environments where privacy, budget, and control are priorities.