When artificial intelligence is mentioned, images of large data centers often appear in people’s minds. Dissatisfaction with AI data centers has become increasingly widespread in recent months because the data centers consume enormous amounts of power and can even heat up their surroundings. The Squeeze Labs team is now proving that AI can also work in a completely different way.
Read more after the ad
An AI with a crank
Their work has the appropriate name “CrankGPT”. From the outside, the AI gadget looks like a red box with a large crank. However, there is enough technology inside to be able to operate artificial intelligence locally and without high power consumption. To do this, the hobbyists use a Raspberry Pi 5 and a crank generator that can produce 20 watts. The generator is actually used to charge devices via USB with muscle power in an emergency.
As soon as the crank is turned, the built-in Raspberry Pi switches on. According to the inventors, it was therefore extremely important to choose a fast-loading operating system. Otherwise you would have to crank for minutes before something happens. They chose DietPi, a minimalist Debian version. This means Linux starts in less than three seconds.
The complete startup process takes around 30 seconds. It takes around ten to 15 seconds for the Raspberry Pi to fully boot. There are also the mentioned three seconds for Linux and ten to fifteen seconds in which the model is loaded. The AI can then be asked questions. CrankGPT is even capable of translating language. Speech recognition is done via Moonshine ASR, while the speech itself is processed by an AI model. The AI’s response is ultimately converted from text to speech via Piper.
For the AI, the hobbyists have tested several models that work reliably. These include Liquid AI LFM 2 in the variants with 350 million and 1.2 billion parameters as well as Gemma 3 with one billion parameters. All of them can produce quick answers without much latency, even though the hardware isn’t exactly bursting with performance. Other models such as Qwen 3.5 2B could only generate single-digit tokens per second – too slow for responses intended to occur in real time.
The hobbyists write about their work: “Even if it is currently not practical to run sophisticated AI workloads on a Raspberry Pi, our work suggests that there is an entire class of undiscovered AI applications that can run locally without consuming large amounts of energy. And because models are becoming smaller and more efficient, at some point they will no longer only run on the current iPhone, but also on smaller and significantly cheaper hardware.”
Read more after the ad
This post is up first t3n.de appeared.
(jl)
