Short answer
- Want the fastest answers from models up to about 30B? A PC with a 24–32 GB NVIDIA card.
- Want to run the biggest models (70B and up) on one quiet machine? A Mac with 64–128 GB.
- Already own one of them? Start there. Both run the popular models well.
Two numbers decide everything
Memory size decides which models fit. Memory bandwidth (how fast the chip can read that memory) decides how fast the model writes. Every word the model writes means reading the whole model once, so speed is roughly bandwidth divided by model size.
Graphics cards have very fast memory but not much of it. Macs share one big pool between everything, so they have far more room, but it is slower to read.
| Machine | Bandwidth | Qwen3.8 27B | Speed | Biggest model that runs well |
|---|---|---|---|---|
| PC with RTX 5090 (32 GB) | 1,792 GB/s | Runs great | 55–90tok/sFast | Qwen3.6 35B-A3B |
| PC with RTX 4090 (24 GB) | 1,008 GB/s | Runs well | 35–50tok/sFast | Qwen3.6 35B-A3B |
| Mac with M5 Max (128 GB) | 614 GB/s | Runs great | 19–35tok/sSmooth | DeepSeek V4 Flash 0731 |
| Mac with M4 Pro (48 GB) | 273 GB/s | Runs great | 9–15tok/sReadable | Kimi Linear 48B-A3B |
Speeds are our estimates from memory bandwidth, shown as a range. See our measurements for real runs.
Other things that matter
- Long prompts: Macs are slower at reading a long prompt before they start answering. With big documents or large code files, the wait before the first word can be long.
- Noise and power: a Mac runs big models quietly at low power. A high-end graphics card can draw 450 W or more under load.
- Software: almost every AI tool supports NVIDIA first. Macs are well supported by Ollama, LM Studio and MLX; image and video tools sometimes lag.
- Mac memory limit: by default macOS lets the graphics side use part of the memory: about two-thirds on a 24–32 GB Mac, rising to about 56 GB of a 64 GB Mac and about 120 GB of a 128 GB Mac. A one-line command raises it; our checker marks these cases as “Fits with a tweak”.
Before you buy
Check the exact model you care about on its model page, then compare real prices on our hardware pages. Prices for memory and graphics cards have moved a lot in 2026.