Colibri Engine Runs 744B AI Model on a Laptop Using Just 25 GB RAM

Colibri is an open-source, pure C inference engine that enables GLM-5.2, a 744-billion-parameter AI model developed by Z.ai, to run on consumer laptops with as little as 25 GB of RAM. Instead of loading the entire model into memory, Colibri streams only the required 'expert' parameters from a local NVMe SSD on demand, keeping resident RAM usage around 10 GB. The model's Mixture-of-Experts architecture means only a small subset of its 21,504 routed experts are needed per token, making disk-based streaming viable. Performance is significantly slower than cloud alternatives — reaching just 0.28 to 1.83 tokens per second versus 30–50 on an H100 GPU — but the setup runs entirely offline with no API costs or data sharing. Colibri supports Linux, macOS, and Windows, and requires approximately 370 GB of local storage for the quantized model weights.