Colibrì proof-of-concept gains frontier-level 1.5-TB AI model (www.tomshardware.com)

Italian engineer Vincenzo, known as JustVugg, has developed a proof-of-concept named Colibrì that enables users to run the 1.5-TB GLM-5.2 AI model, comprising 744 billion parameters, on a modest home setup using just 25 GB of RAM and a virtual NVMe drive. This project emerges amidst rising subscription costs for AI bots and mounting data privacy concerns, offering a novel solution for enthusiasts who wish to explore advanced AI models without the hefty price of specialized hardware like Nvidia’s NVL72 rack. While Colibrì operates at a slow pace of 0.05 to 0.1 tokens per second—a serious limitation for real-time applications—it leverages a Mixture-of-Experts architecture, where hundreds of experts are activated on a per-token basis to deliver high-quality responses. Colibrì’s innovative approach allows it to dynamically load and unload expert models based on user queries, making it feasible to utilize high-capacity models on less powerful machines. However, the model's performance is heavily constrained by storage I/O and memory bandwidth, requiring ongoing adjustments and optimizations. As this project gains traction and evolves, it holds the potential to make high-level AI capabilities more accessible to the average user, paving the way for a new era of affordable, advanced AI experimentation in home labs.