Buying guide
The best board for local LLMs.
Short answer, by budget, then the numbers behind it. Every figure is a measured, sourced run from our database. No estimates.
Updated 2026-10-17. Every speed links to its source; every price is sourced and dated.
The short answer.
- Under $150: M5Stack LLM Module Kit (≈ $79.90 for the board plus the parts it needs to run). See the full kit Check price on Amazon
- Under $250: Radxa ROCK 5B+ 16GB kit (≈ $198.79 for the board plus the parts it needs to run). See the full kit Check price on Amazon
- Under $500: Jetson Orin Nano Super kit (≈ $449.99 for the board plus the parts it needs to run). See the full kit Buy at Seeed StudioBuy on Amazon
The numbers.
| Board | Llama 3.2 1B | Qwen2.5 1.5B | Llama 3.2 3B | Llama 3.1 8B |
|---|---|---|---|---|
| Raspberry Pi 5 from $175 | 11.22 | 11.73 | 5.95 | 2.05 |
| Raspberry Pi 5 + AI HAT+ 2 from $180 | 8.7 | 6.82 | 2.65 | – |
| Rockchip RK3588 from $135 | – | 16.69 | – | 3.72 |
| Jetson Orin Nano Super from $399 | 10.96 | 9.37 | 43.1 | 19.1 |
| Jetson AGX Orin from $3,499 | 163.9 | – | 80.4 | – |
| M5Stack Module LLM from $79.90 | 4.48 | 4.63 | – | – |
Best sourced generation speed in tok/s; click a number for every run and its source. "–" means no sourced run yet. Prices are listed board prices from our price list.
How to choose.
- Small models (1–2B) are fine on a Pi 5. Its best sourced run on Qwen2.5 1.5B is 11.73 tok/s (source ↗), above the 10 tok/s mark we treat as comfortable chat speed. An RK3588 board reached 16.69 tok/s on the same model using its NPU (source ↗), but needs RKLLM-converted models.
- 3B models are where the Jetson pulls away. On Llama 3.2 3B the Orin Nano Super's best is 43.1 tok/s (source ↗) against 5.95 on a Pi 5 (source ↗) and 2.65 on the AI HAT+ 2 (source ↗).
- 8B models need a GPU or a lot of patience. Llama 3.1 8B: 19.1 tok/s on the Orin Nano Super (source ↗), 3.72 on RK3588 (source ↗), 2.05 on a Pi 5 (source ↗).
- Count the extras. A Pi 5 needs a power supply, a cooler and a microSD card; the Jetson includes its power supply. Each board page has a "What you need to get running" list, and the tokens-per-dollar ranking shows which board gives the most speed per dollar.