What hardware do you actually need?
Pick a model size and a target tokens-per-second. We'll show which hosts have a published figure that meets the spec, and the cheapest of them. Where nobody has published a figure, the table says so: read the methodology below before you buy.
Published measurements only. Tokens per second come from published tests with their setup: the Slips project for the Raspberry Pi 5 and llama.cpp's Apple Silicon scoreboard for the Mac Mini M4. “No figure” means nobody has published one, not that the host cannot do it. Prices and watts are orders of magnitude. Measure before you buy.
| Host | Capex | RAM | tok/s | basis | watts | verdict |
|---|---|---|---|---|---|---|
| Mac Mini M4 16 GB | €749 | 16 GB | 24 | published | 16 | meets target |
| Raspberry Pi 5 8 GB | €138 | 8 GB | no figure | 11 | no figure | |
| Intel NUC 13 i7 | €620 | 32 GB | no figure | 28 | no figure | |
| Single-GPU box (RTX 4060 Ti 16 GB) | €1180 | 32 GB | no figure | 91 | no figure |
Methodology
Until October 2026 this page said its throughput numbers came from a 21-day benchmark of ours. We have no data to show for that benchmark, so the claim has been removed, and the estimates that replaced it went later the same month: they had no measurement behind them either. The Raspberry Pi 5 figures come from the Slips project's tests (Ollama, Q4_K_M: about 11 tok/s on 1B models, about 5 on 3B models; no 7B figure). The Mac Mini M4 figure comes from llama.cpp's Apple Silicon scoreboard (24.11 tok/s generating with Llama 2 7B at Q4_0 on the M4 with the 10-core GPU, read on 17 October 2026). Nobody has published figures for the other sizes and hosts with their setup, so the table shows “no figure”; measure on the hardware you are considering. Prices and watts are orders of magnitude. See GPU vs CPU: when each wins for the reasoning. All figures are for 4-bit quantisation.