CrankGPT looks like a joke website. The hand-cranked AI inside it is real
The landing page is so absurd that half the thread couldn't tell if it was satire. Behind it: a Raspberry Pi 5 running a local LLM, powered entirely by a hand-crank generator. The crank gets harder to turn while the model thinks.
A 20W emergency hand-crank charger feeds a Pi 5 through a custom capacitor board that buffers about 20 seconds of power. A 350M model hits ~49 tokens/sec; idle draw is 4W, and an answer-plus-speech burst spikes toward 5A.
You can feel the inference load in your wrist: when the model and text-to-speech run together, the crank physically stiffens. Half the thread wanted to buy one; the other half pictured humans on hamster wheels for the AI overlords.
Yes, it's real, and you power it by hand
CrankGPT's website is deliberately ridiculous enough that the thread genuinely argued about whether it existed — "I watched the 'is this real' video and I'm still not sure if it's real lol." The technical documentation settles it: this is a working offline voice assistant powered entirely by a hand crank.
The build is a Raspberry Pi 5 (8GB) running speech recognition, a small LLM, and text-to-speech on CPU only, fed by "a cheap off-the-shelf switchable voltage 20W hand-crank generator marketed for emergency USB charging." Because inference makes the voltage sag, the builders added a capacitor board buffering roughly 20 seconds of power. Performance on small models is the surprising part:
| Model | Tokens/sec on Pi 5 |
|---|---|
| LFM2.5 350M | ~49 |
| LFM2.5 1.2B | ~15 |
| Gemma 3 1B | ~14 |
Power draw runs from ~4W idle to ~15W when the model and speech synthesis fire together, with brief 5A spikes — which you feel directly: "when LLM inference and speech synthesis run together, the crank gets a lot harder to turn."
The thread couldn't decide whether to laugh or order one
The comedy and the sincerity arrived together. One reader, target demographic identified, riffed about wiring a rowing machine to power a laptop answering prompts. Another supplied the dystopian caption: "In the future, humans will just all be on hamster wheels generating power for our AI overlords." And a straightforward buyer cut through it: "TBH I'd buy that."
The most useful comment did the navigation the site refused to: the absurd front page hides a link to documentation of an actual working hand-crank Raspberry Pi running a local model — the joke is the wrapper, the engineering is the gift.
A toy that makes a real point
Strip the bit and CrankGPT is a tidy demonstration: a useful conversational AI now fits inside a CPU-only single-board computer and a power budget a human arm can supply. That's a statement about how far small models have come — a 350M model at 49 tokens/sec from muscle power. The website wants you to laugh. The tokens-per-crank table wants you to notice that the floor for "runs an LLM" has dropped to a literal emergency charger.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Why counting CVEs to compare Rust and C++ is a trap, in one curl example
Pass NULL to a C function, get a segfault, and it's 'wrong usage,' not a CVE. Do the equivalent in safe Rust and any memory bug is the library's fault by definition. The post argues that's why raw CVE counts mislead.