CrankGPT Is Guaranteed to Make You Cranky

It’s hard to have a conversation about AI these days without the topic of energy consumption coming up. The latest wave of large language models and image generators is far from efficient in this regard. There is hope that more efficient algorithms will be developed in the future, but until that time, Squeez Labs has cranked out a project that could help. In truth, it won’t really help at all, and using it will only make you cranky. Even so, it’s a cool project that you won’t want to miss.

The team has built a device they call CrankGPT, and as you might have guessed from the name, it is a device running a large language model that is powered entirely by a hand crank. Unless you've hooked that crank up to a gas engine, you’re not going to generate enough electricity to power a modern GPU. But Squeez Labs did manage to keep a Raspberry Pi 5 powered up long enough to boot and act as an AI voice assistant with nothing but manual labor.

Spinning up a prototype

The system is built around a stock Raspberry Pi 5 equipped with 8 GB of RAM and a cooling fan. The computer is paired with a Seeed Studio ReSpeaker microphone HAT that provides both audio input and output, allowing users to interact with the system entirely through voice. The real challenge, however, was not getting the hardware to run AI models locally. It was keeping the computer alive while powered by a hand-crank generator.

A Raspberry Pi running AI workloads is quite demanding. Speech recognition can push power consumption to around 8 watts, while running a language model and text-to-speech engine simultaneously can require roughly 15 watts. The team found that current spikes during inference were enough to trigger the generator's protection circuitry and cause brownouts. To solve this problem, they designed a custom capacitor board that acts as a temporary energy reservoir, smoothing voltage fluctuations and providing enough stored energy to keep the system operating during brief interruptions in cranking.

Rather than relying on cloud services, every component runs locally on the Pi. A stripped-down DietPi installation helps minimize boot times, while the voice assistant stack was written specifically for low-latency operation on modest hardware. Speech recognition is handled by Moonshine ASR, voice activity detection comes from Silero VAD, and responses are generated by smaller language models running through llama.cpp.

Larger models were simply too slow to create a natural conversational experience. Instead, the developers settled on smaller options like Liquid AI's LFM2 models and Gemma 3 1B. The fastest configuration can begin responding in under a second, which is impressive for a fully offline voice assistant running on a single-board computer.

Text-to-speech duties fall to Piper, which was chosen because it could synthesize speech significantly faster than competing lightweight solutions. Responses are streamed sentence-by-sentence, allowing the device to begin speaking before the entire answer has been generated.

CrankGPT won’t solve the world’s energy problems, but if you ever find yourself without power you can still get your AI fix.