Raspberry Pi AI

1. Buy a Raspberry Pi 5 (8 GB, $80) + 32 GB microSD ($10) + power supply (~$15).
2. Flash Raspberry Pi OS (64-bit, Bookworm) via Raspberry Pi Imager.
3. Install Ollama for ARM64: curl -fsSL https://ollama.com/install.sh | sh.
4. ollama pull llama3.2:1b (700 MB — runs at 5-10 tok/s on Pi 5 CPU) or ollama pull tinyllama (600 MB, 8-15 tok/s).
5. ollama run llama3.2:1b "Explain how a transistor works in simple terms." — first response in 10-30 seconds.
6. For vision: pip install ultralytics → YOLO11n runs at 3-5 fps on Pi 5 CPU (no GPU). With a Hailo-8L accelerator (~$70, PCIe HAT), YOLO runs at 30+ fps.
7. For audio: Silero VAD + Whisper tiny on Pi 5 CPU — VAD in real-time, Whisper transcodes at ~0.5× real-time.
8. Use cases: home automation AI, offline voice assistant, DIY security camera, edge data logger with anomaly detection.

Raspberry Pi 5 8 GB (~$80) is the AI entry point for makers. Runs Llama 3.2 1B at 5-10 tok/s, YOLO11n at 3-5 fps, Whisper tiny at 0.5× real-time — all on CPU at 15W. For a home AI project (voice-controlled automation, person-detecting doorbell, plant health monitor): total hardware $100-150 (Pi + camera + sensors). The Pi 5's BCM2712 CPU has ARMv8.2 vector extensions that accelerate quantized inference ~2-3× vs. Pi 4. For $150, you get a fully functional AI device that runs 24/7 on $5/year of electricity. Pi AI is for learning, prototyping, and single-purpose always-on devices.

Raspberry Pi 5 8 GB ($80) + Hailo-8L AI accelerator HAT ($70) = $150 for real-time AI. The Hailo-8L delivers 13 TOPS at 5W — YOLO11n at 30+ fps, ResNet-50 at 100+ fps, SSD at 30+ fps. For a production edge device (retail footfall counter, factory defect detector, agricultural crop monitor): Pi 5 + Hailo-8L + industrial camera = ~$250 per node. For LLM-capable edge: Pi 5 runs 1B-3B models on CPU acceptably for non-real-time use. For real-time LLM inference at the edge, upgrade to Jetson Orin Nano ($500, see /hardware/jetson-ai). Pi AI is for when $/Watt matters more than raw performance. A fleet of 100 Pi+Hailo nodes ($25K) does 100 concurrent real-time vision streams at 1,500W total — equivalent to ~2 RTX 4090s in terms of concurrent stream count (but not per-stream quality).

The mistake: Buying a Raspberry Pi 5 4 GB model (instead of 8 GB), installing Ollama, pulling a 7B model, and wondering why the Pi crashes or runs at 0.1 tok/s. Why it fails: The Pi 5 shares RAM between CPU and GPU. A 7B Q4 model needs ~4.5 GB for weights alone. On a 4 GB Pi, with 1 GB for the OS, there's 3 GB available — the model doesn't fit. The Pi starts swapping to the SD card (100× slower than RAM), dropping inference to effectively zero. The fix: Buy the 8 GB Pi 5 for AI. Run models sized for the available RAM: 1B-3B models on 8 GB Pi (3B Q4 = ~2 GB weights + 2 GB OS/overhead = 4 GB, fits). For 7B models, you need a device with 12+ GB RAM (Jetson Orin, x86 mini PC). Match model size to available RAM. On a Pi with 8 GB, usable model max is ~3B parameters at Q4 quant. The Pi is not a desktop GPU — it's a low-power edge device. Size models accordingly.