what is colibri fr?

Colibri is a C program that lets you run GLM-5.2, a huge 744 billion parameter AI language model, on an ordinary consumer computer with around 25 gigabytes of RAM, something that would normally require a datacenter's worth of memory. It manages this by keeping only the small part of the model that is always needed loaded in RAM, while streaming the rest from your hard drive or SSD only when it is actually needed for the current word being generated. The trick that makes this possible is how this type of AI model works internally. A model built this way, called a mixture of experts, only activates a small fraction of its total parameters for any single word it generates, roughly forty billion out of seven hundred forty four billion, and only about eleven gigabytes worth of those active parameters actually change from word to word. Colibri keeps the always used core of the model resident in memory, while the thousands of specialized expert components live on disk and get pulled in on demand, using a cache that learns which experts get used often and keeps those ready. The project compares this approach to how a just in time compiler works: rather than loading everything up front, it watches what is actually needed and stages that data just in time. The engine itself is described as a single C file with a few small header files, requiring no BLAS math library, no Python at runtime, and no GPU, though a GPU can be used if available to hold more experts in faster memory for better speed. If you have a second SSD available, the tool can split reads of the model across both drives at once to increase how fast experts can be streamed in, with the model mirrored across both and falling back gracefully if one drive is unavailable. The same program scales from a modest twenty five gigabyte laptop, where nearly everything streams from disk, up to larger machines where the whole model fits in memory and disk is not needed at all during normal use. It ships with a web dashboard that shows live metrics about which experts are firing, their storage tier, and overall performance.