You stand in a tarp-lit tent, rain slapping the canvas, and a hand crank is the only thing between you and an answer. I turned the handle, felt resistance, and the gadget spoke back in a flat, calm voice. For a second the apocalypse stopped feeling inevitable and started to feel manageable.
At a cramped kitchen table I first tested a hand crank that powers machine intelligence
You and I both remember the jokes about solar panels and generators in survival forums. CrankGPT is the practical rejoinder: a battery-less, hand-cranked box built to run a local large language model when the grid is gone. It’s an experimental proof of principle from SqueezLabs that shrinks an LLM setup down to a handful of parts you could stash in a bug-out bag.
The hardware list reads like a maker’s manifesto: a Raspberry Pi 5 with 8GB RAM, an audio input/output card, a custom capacitor board, and a 20W hand generator. The team wired the generator to their board so the Pi sees steady voltage, and they left a human in the loop—you literally feel computation through the crank: inference and speech synthesis make the handle tighten under your hand.
What is CrankGPT?
CrankGPT is a self-contained conversational system that takes voice input, runs a small LLM locally, and speaks back using open-source TTS. SqueezLabs designed it to run models such as Liquid AI’s LFM2 variants (350M or 1.2B) or Gemma 3 at 1B, and it uses Piper for text-to-speech with a custom voice agent converting speech to tokens. It’s not a replacement for cloud-scale systems; it’s a fallback—an offline advisor when connectivity is gone.
At a maker fair someone handed me the crank and challenged my stamina
Running inference and TTS together is surprisingly demanding, so the designers used smaller models to keep power and token counts low. SqueezLabs’ ethos was simple: don’t burn kilowatts when a compact model will do the job. CrankGPT is a pocket-sized power station for conversational AI—small, deliberate, and designed around human effort rather than wall sockets.
That choice also points at a larger industry tension. You’ve seen OpenAI and other companies push bigger models with massive compute costs. What SqueezLabs shows is a counter-approach: precision over brawn, where you trade raw scale for efficiency, privacy, and offline resilience. If you value being able to ask questions when every cell tower is silent, smaller models matter.
How does CrankGPT work without mains power?
The generator produces up to 20W into a capacitor-backed regulator that smooths supply to the Raspberry Pi. Speech arrives through the audio card, is turned into text by the custom voice agent, fed to a local model, and then routed to Piper for speech output. You literally crank to allocate compute; when the model works harder, the handle resists harder—the mechanical feedback is part UX, part battery indicator.
At my bench I swapped models to see accuracy differences
Using a 350M Liquid LFM2 yields snappy responses and much lower crank effort; moving to a 1.2B or Gemma 1B improves nuance but demands more torque. The trade-off is plain: richer language needs more cycles and more sweat. This is where you decide whether you want a terse survival checklist or a full conversation about cooking foraged roots.
There’s also the privacy angle: running everything locally means no telemetry leaving your tent. For journalists, field researchers, and anyone wary of cloud logging, that matters. If you’re suspicious of handing your emergency plans to a corporate server, CrankGPT keeps the dialogue on-device.
At midnight I asked it what to eat from a half-empty pantry
The responses were practical and sometimes oddly human—recipes from canned beans and stale bread, notes about boiling water, warnings about unknown mushrooms. It doesn’t hallucinate with the polished confidence of a cloud model, but it keeps you out of the most dangerous mistakes. CrankGPT is a stubborn mule under load: it won’t wow you with encyclopedic recall, but it won’t keel over when you need it most.
There are limits. You won’t get the run of GPT-4-style reasoning, and updates require physical access to the device and model files. The team’s point about energy footprint is worth repeating: a thousand-token cloud call burns far more energy than a local 350M inference, and that matters if your power is your own two hands.
Tools in this space tie into recognizable names: Raspberry Pi hardware, Liquid AI’s foundation models, Gemma 3, Piper for speech, and the maker ecosystem that surrounds open-source AI. Think of CrankGPT as an intersection of those projects—practical engineering, not smoke-and-mirrors marketing.
If you’re imagining a future where every question must route through a corporate API, this project offers a different mental model: conversational intelligence you can carry, power by hand, and control on-device. That trade-off will appeal to tinkerers, privacy-minded users, and anyone who plans to be offline for extended stretches.
Would you trust a hand crank to answer life-or-death questions when the servers are down?