Colibrì is a pure-C, zero-dependency inference engine that lets large Mixture-of-Experts models run on ordinary consumer hardware by treating VRAM, RAM, and disk storage as one unified memory hierarchy. Rather than loading an entire model into memory, it keeps only the dense layers resident in RAM and streams the routed experts from disk on demand based on the model’s routing patterns. The engine currently supports four model families, including GLM-5.2 at 744 billion parameters, and scales from about 0.05 tokens per second on resource-constrained machines up to 6.8 tokens per second on high-end GPU setups using the same codebase.
