Colibrì proof-of-concept gains frontier-level 1.5-TB AI model — novel approach runs on only 25GB of RAM and shows promise for local AI setups

Italian engineer Vincenzo built Colibrì, a proof-of-concept tool that runs the massive 1.5-TB GLM-5.2 AI model on just 25GB of RAM and a basic CPU. It works by loading small pieces of the model into memory one at a time, rather than needing expensive datacenter hardware. The tradeoff is extremely slow speed — around 0.05 tokens per second — but the response quality is reportedly impressive, showing real promise for future local AI setups.