Open-Source Colibri Engine Runs 744-Billion-Parameter Models on Consumer Hardware Using Disk Streaming

A pure-C inference engine called Colibri, created by developer Vincenzo Fornaro and released in July 2026, enables frontier-scale mixture-of-experts language models to run on consumer laptops and desktops by treating storage, system RAM, and VRAM as a unified memory hierarchy. The breakthrough exploits a fundamental asymmetry in Mixture-of-Experts (MoE) models: while containing hundreds of billions of parameters, these models activate only a small fraction per token, allowing inactive experts to remain on fast NVMe storage and be streamed on-demand rather than loaded entirely into memory.