AI on the Raspberry Pi 5: LLM-Powered Arcades and Agentic Assistants
AI is moving to the edge — and the Raspberry Pi 5 is right at the front of it. Inspired by two projects in Raspberry Pi Official Magazine #166, here's how makers are running large language models and autonomous AI agents on a Pi, and why the new 16GB Raspberry Pi 5 we now stock is the ideal board to do it on.
Artificial intelligence used to mean racks of GPUs in a distant data centre. Not any more. A new wave of projects is running large language models (LLMs) and even autonomous AI agents on a board that fits in your hand — the Raspberry Pi 5. The June 2026 issue of Raspberry Pi Official Magazine (#166) featured two standout examples, and both point to the same conclusion: with enough RAM, a Pi 5 makes a genuinely capable little AI machine.
Why the Raspberry Pi 5 — and why 16GB — for AI
When you run AI workloads, memory is the headroom. Bigger models, longer context, more background tools and more of the operating system kept in RAM all compete for the same space. Both magazine projects call out RAM as the thing that matters most: the OpenClaw agent tutorial recommends a Pi 5 with 8GB RAM or higher, and the AI Arcade console was built on an 8GB Pi 4.
That's exactly why the model we stock is the 16GB Raspberry Pi 5 (BCM2712) — the top of the Pi 5 range. The extra memory is the difference between just fitting a model and having room to run it comfortably alongside a browser, a local web app, an agent's tool stack, or a local model on an AI accelerator HAT.
The board behind these projects — in stock now
Project 1: AI Arcade — a console that invents its own games
London developer Grigor Todorov built AI Arcade: a retro games console that never runs out of games, because it generates them on the fly with an LLM. Power it on and the Pi boots straight into a local web app in full-screen Chromium; from there it calls a cloud LLM (via the OpenAI API) to invent a brand-new mini-game each time you play.
The clever part is how lightweight it is. “First, the app asks the model to generate four questions, each with two possible answers,” Grigor explains. Your choices feed a second prompt that returns a complete HTML game, which is displayed and played with the arcade buttons. Two LLM calls, and you have a playable game in seconds. The whole app was ‘vibe-coded’ — described in plain language to an AI rather than written by hand.
“The system is able to produce simple, addictive games that evoke the nostalgia of a bygone era.”
Grigor built it on an 8GB Pi 4 and notes the weakest part is the visuals — exactly the kind of richer, higher-quality generation that benefits from the extra compute and memory headroom of a 16GB Pi 5.
Project 2: OpenClaw — an agentic AI assistant on your Pi
The second project goes further than generation — into autonomy. Magazine editor Lucy Hattersley's tutorial sets up OpenClaw, an agentic AI assistant, on a Raspberry Pi 5. Unlike a chatbot that just answers questions, an agent takes actions: with your permission it can read your email and calendar, create and edit files, browse the web, send messages, and — thanks to the Pi's GPIO pins — control real hardware and home automation.
The tutorial's hardware shopping list is a good blueprint for any serious AI Pi build: a Raspberry Pi 5 with 8GB RAM or higher, a 128GB NVMe drive (agents write to storage constantly, so skip the microSD card), a power supply, and an OpenAI API key. The Pi 5 is described as the ideal platform precisely because it gives the agent an isolated, contained environment to work in — a sandbox you control rather than turning it loose on your main computer.
“We have an agentic coder running on our Raspberry Pi, ready to do all kinds of weird and wonderful things.”
Two ways to run AI on a Raspberry Pi 5
Both projects start with a cloud LLM (the OpenAI API) because it's the quickest way to get going. But the Pi 5 can also run a model locally on an AI accelerator HAT — no internet, no per-token bill, and your data never leaves the board. Here's how the two approaches compare:
Cloud LLM vs local LLM on a Pi 5
| Feature | Cloud LLM (OpenAI API) | Local LLM (AI accelerator HAT) |
|---|---|---|
| Where the model runs | A data centre | On the Pi itself |
| Internet needed | Yes | No — runs offline |
| Ongoing cost | Pay per token (API) | None after hardware |
| Privacy | Data sent to the provider | Data stays on your Pi |
| Setup effort | Lowest — just an API key | More setup; needs a HAT |
| Best for | Getting started fast, biggest models | Privacy, no running cost, edge deployment |
The more RAM your Pi has, the more capable the local option becomes — another reason the 16GB board is the one to build on.
What you'll need to get started
If either project has you reaching for a soldering iron (or just an SSH session), here's the core kit:
- Raspberry Pi 5 — 16GB recommended for AI headroom
- NVMe SSD (e.g. 128GB) on an M.2 HAT or a USB 3.0 enclosure — agents and local models write to storage constantly, so avoid microSD
- Official Raspberry Pi 5 power supply — stable power matters under load
- An LLM — an OpenAI API key to start, or a local model on an AI accelerator HAT for offline, no-cost inference
- A fresh install of Raspberry Pi OS (Raspberry Pi Imager), plus an internet connection for the initial setup
Want to build your own AI projects? The 16GB Raspberry Pi 5 is the board these makers reach for — and we have it in stock now.
Shop the Raspberry Pi 5 16GBProjects and images featured in Raspberry Pi Official Magazine #166 (June 2026): ‘AI Arcade’ by David Crookes (maker: Grigor Todorov) and ‘Set up OpenClaw on Raspberry Pi’ by Lucy Hattersley.
Some links in our articles may be affiliate links — if you buy through them we may earn a small commission at no extra cost to you. See our Affiliate Disclosure.
