Find the right local AI machine.
Choose the model you want to run, set a budget, and compare fixed hardware configurations instead of broad memory ranges.
| Hardware | Vendor | Memory | Cost | 4-bit | 5-bit | 8-bit | None | Bandwidth | TFLOPs | Fit |
|---|
Planning estimates only. Each card uses a representative fixed configuration so the comparison is easier to understand. Real prices, speed, memory overhead, usable context length, model architecture, and software support will vary. 5-bit and 8-bit maximum model sizes are interpolated between the listed none and 4-bit figures for rough planning. Actual stack support varies.
Photos are illustrative: many are CC-licensed images from Wikimedia Commons (similar model or product family, not always the exact SKU). NVIDIA RTX 5090 cards use RTX 4090 imagery until widely published press photos exist; Intel Arc entries use an Arc B580 Limited Edition shot as a stand-in for B60-class hardware. GB10 / DGX Spark uses a Commons screenshot of Project DIGITS with the GB10 Grace Blackwell superchip (Geekerwan), as described on the file page, which aligns with the Nvidia DGX / DGX Spark section on Wikipedia.