CrankGPT Runs Offline AI on a Raspberry Pi 5 You Power by Hand
A maker project called CrankGPT runs a fully offline AI voice assistant on a Raspberry Pi 5 powered entirely by a hand crank, with the crank getting harder as the model thinks.
The Most Delightful Edge-AI Project of the Month
Sometimes a build comes along that is less about raw specs and more about pure ingenuity and joy. CrankGPT, documented by Squeez Labs and picked up by Hackaday on June 11, 2026, is exactly that: a fully offline AI assistant running on a Raspberry Pi 5 that you power entirely by turning a hand crank. No wall outlet, no battery to charge overnight — just human elbow grease converted directly into tokens.
How the Build Works
The hardware is refreshingly approachable. It starts with a stock Raspberry Pi 5 (8GB), topped with a cooling-fan HAT and a voice-assistant I/O HAT for microphone and speaker duties. Power comes from a hand-crank USB charger, and because cranking produces uneven bursts of electricity, the builder added a custom capacitor board to smooth out the spikes and keep the Pi from browning out mid-sentence. It is a clean illustration of a classic single-board computer principle: the Pi itself is sober and capable, and the magic is in the HATs and power conditioning stacked around it.
On the software side, CrankGPT runs llama.cpp with a thoughtfully chosen set of small models: Liquid AI's LFM2 in 350M and 1.2B sizes, plus Google Gemma 3 (1B). Those compact models are the whole point — they are small enough to run responsively on the Pi's ARM processor without a network connection, handling voice interaction, image tasks, short poems, and even bits of code entirely on-device.
The Detail That Makes It Sing
Here is the flourish that has the maker community grinning: as the model generates a response, the crank physically gets harder to turn. More computation means more power draw, and you feel it in your arm. It is an almost poetic, tactile reminder of the energy cost of inference — something we usually never sense when a data center does the work for us. For an edge-AI project, it is a brilliant teaching device disguised as a gadget.
Why Projects Like This Matter
It would be easy to file CrankGPT under "fun hack and move on," but I think it captures something important about where compact computing is heading. The fact that a hand-cranked Raspberry Pi can run a useful offline assistant at all is a testament to how far small, efficient models and capable single-board computers have come. Fully local, no-connectivity AI is genuinely useful — for privacy, for resilience, and for places without reliable power or internet.
CrankGPT is also just the kind of accessible, reproducible maker project that inspires people to try their own builds. The parts are common, the software is open, and the concept is irresistible. If you have a spare Pi 5 and a hand crank lying around, you now know what to do this weekend.
Sources: Hackaday — "AI, the truly environmentally friendly way," June 11, 2026; Hackster.io — CrankGPT project coverage, June 2026; Squeez Labs project page (squeezlabs.github.io/handcrank).