2026-10-06

Raspberry Pi has published a detailed walkthrough showing how to run CNN, VLM (vision-language models) and SLM (small language models) entirely on a Pi 5 — no cloud, no GPU server, just the £80-ish board and its NPU/CPU combo.

The official blog post walks through a pipeline where a camera feed is classified by a lightweight CNN, then a VLM describes what it sees, and finally an SLM generates a natural-language response. All local, all low-power.

Why homelabbers should care

If you've been eyeing local AI but can't justify a £500+ GPU, this is genuinely interesting. A Pi 5 draws around 5–10W under load. Running a vision-and-language stack on that kind of power budget opens up practical projects: smart CCTV summaries, doorbell descriptions for accessibility, or feeding context into Home Assistant automations.

The models demonstrated are small — don't expect GPT-4-level reasoning. But for structured, narrow tasks ("is there a person at the door?" → "yes, they're carrying a parcel" → "send a notification") this is more than enough. The Pi foundation clearly wants to position the Pi 5 as a genuine edge-AI platform, not just a hobbyist toy.

Honest caveats

Inference is slow compared to even a modest desktop GPU. You're looking at seconds per response, not milliseconds. Model choice is limited to what fits in 8GB RAM. And the software stack still requires some fiddling — this isn't a one-click install yet.

What to do next

Grab a Raspberry Pi 5 8GB if you haven't already — the 8GB model is essential for these workloads. You'll also want a decent micro SD or NVMe hat for model storage. If you want to go further and build agentic AI workflows around your homelab, our The Homelab Agent Kit — 9 AI Agents, 13-Chapter Build Guide, Python walks you through exactly that with Python and practical build guides.

Worth bookmarking the source post and experimenting this weekend.