A developer based in Italy has successfully run a 744 billion parameter AI model named Calibri on a laptop lacking a dedicated graphics processing unit.
The project is built on GLM 5.2, an open model described as capable of outperforming paid alternatives.
Calibri works by keeping only around 10 gigabytes of the model in the computer's RAM at any given time.
The remainder of the model streams from the hard drive as needed, in a manner comparable to how Netflix streams video content rather than downloading an entire film before playback begins.
This streaming approach challenges a long-standing industry assumption that expensive GPUs are strictly necessary to run models of this scale.
Getting it running
The installation process is designed to be straightforward.
Users clone the project's repository and run a single setup file, which then builds the project automatically.
The main storage requirement is the model download itself, which occupies 370 gigabytes of disk space.
Using the model day to day
Once installed, users can interact with the model offline simply by typing "Kaylee" into a chat interface.
Developers can also run the command "Kaylee serve" to turn the model into an OpenAI-compatible endpoint.
Changing a single line of code to point toward that local endpoint lets developers eliminate token costs entirely, since all data processing then remains on the local machine rather than being sent to an external server.
Where to find it
A GitHub repository containing the full installation guide is available for anyone wanting to try the setup themselves.
