There is a particular joy in making a small model do something surprisingly useful. It feels less like renting a supercomputer and more like teaching a curious little machine a new trick.
At ajolote.io, we have been testing local models in the spaces between client work: on laptops, on tiny servers, and in conversations where latency is part of the experience.
The lesson so far is simple: the best model is often the one that fits the shape of the problem. A little context, a clear job, and permission to be weird goes a long way.