DoomReady · Guide
Can a Raspberry Pi run AI offline? Measured speeds and what's usable
Yes. A Raspberry Pi 5 with 8GB of memory runs small language models entirely offline. DoomReady measured about 11 tokens per second on a 1B model, about 5.2 on Qwen 2.5 3B (roughly reading speed) and about 1.6 on an 8B model, drawing around 11W while answering. The 3B model is the practical sweet spot.
Measured on a Pi 5 (November 2026)
- 1B model, quick answers: about 11 tokens/sec.
- Qwen 2.5 3B, the working model: about 5.2 tokens/sec.
- 8B model, deep mode: about 1.6 tokens/sec (slow, but it works).
- Power: about 3W idle, 6.5W working, 11W while answering.
What it's good for, and what it isn't
Small models are good explainers: summarising an article, putting a medical page in plain words, walking through a repair. They are not reliable on facts by themselves, and they can be confidently wrong.
That's why the Offline Brain pairs the AI with the full offline library: the library is the authority, and the AI is the explainer. Check every important answer against the source page.
How to run one
- Use a Pi 5 with 8GB of memory and fast storage.
- Install a local model runner such as Ollama or llama.cpp.
- Download a small instruction model (1B to 3B) while you still have internet.
- Test it fully offline before you rely on it.
Questions people ask
Does it need the internet?
Only to download the model once. After that it runs entirely on the Pi.
Which model should I use?
A 3B instruction model is the best balance of speed and quality on a Pi 5. 1B is faster but weaker; 8B is better but slow.
Can I run it from a battery?
Yes. At a few watts, a pocket power bank runs it for hours and a small solar panel can keep it going.