Which Pi for What
The question isn't "which Pi is best" — it's "which Pi fits this job." Here's the honest matching, including where a Pi is the wrong tool entirely.
The quick answers
| Job | Pick | Why |
|---|---|---|
| Home server / NAS | Pi 5 4–8GB + NVMe HAT | PCIe kills the SD bottleneck; ~400 MB/s over GigE |
| Budget home server | Pi 4 4GB + USB SSD | Half the price, fine for files and media at 1080p |
| Pi-hole / DNS / ad-blocking | Zero 2 W | Sips power, runs 24/7 for pennies |
| MQTT broker, tiny services | Zero 2 W | 512MB is plenty; ~0.4W idle |
| Small-AI / Ollama tinkering | Pi 4 4–8GB (or Pi 5 8GB) | See the honest notes below |
| Desktop replacement | Pi 500+ (or Pi 5 8GB) | 16GB + NVMe is the only true daily-driver |
| Learning Linux | Pi 4 2GB or Pi 5 2GB | Cheapest full Linux computer |
| Sensors, buttons, LEDs | Zero 2 W (Linux) or Pico 2 W (real-time) | Linux+GPIO vs microcontroller — pick by the job |
| Camera projects | Pi 5 | Dual camera ports, CPU to process frames |
| Retro gaming | Pi 5 | The extra CPU headroom matters for emulation |
| K8s cluster learning | 3–4× Pi 4 or 5 | One per node; the classic use case |
Small AI on a Pi: the honest version
This is where hype outruns reality, so here's what actually happens, measured on a Pi 4 (4GB):
- 1–2B models work (qwen3:1.7b, smollm2:1.7b) at a few tokens/second. Fine for experiments, duels, and learning — not for chat you enjoy.
- Two models loaded at once OOM'd the 4GB board. The fix
was unloading between uses (
OLLAMA_KEEP_ALIVE=0), at the cost of slow model swaps. - It runs hot. Sustained inference pushed past 80°C and into thermal throttling — active cooling isn't optional for this workload.
- The Pi 5 is meaningfully faster (~2× on CPU-bound inference) but still a toy next to any used GPU box.
If your goal is learning how local LLMs work, a Pi 4 is a great cheap lab. If your goal is using LLMs daily, put the money toward the AI-PC build instead (see the Local AI guide).
When it's not a Pi job
- Needs >8GB RAM or real speed — a used mini-PC (Intel N100-class) often costs the same as a kitted-out Pi 5 and runs circles around it. The Pi wins on power draw and GPIO, not raw compute per dollar.
- Battery-powered sensing — a Pico sipping milliwatts beats any Linux Pi. Booting Linux to blink an LED is burning money.
- Real-time control — Linux isn't real-time. Motors, precise timing, bit-banging protocols: Pico.
- Serious NAS with redundancy — one USB disk on a Pi is a file server, not a NAS. For RAID and real data safety, that's the ZFS conversation on bigger hardware.
The fleet angle
Pis make excellent infrastructure precisely because they're cheap and boring: one Pi 4 as a fleet monitor watching your other machines, a Zero 2 W as a serial console server, another as the MQTT broker everything reports to. At $15–60 each, dedicating a whole computer to one job stops feeling extravagant — and a dead-simple box per job is easier to reason about than one big server doing twelve things.