A downloaded model, all on its own, does absolutely nothing. It's billions of frozen numbers in a file — a puppet lying limp on the floor. Something has to pick it up and work it.
That something is the harness — the software that actually runs the model. llama.cpp, Ollama, vLLM. It's the control rig and the hands that bring the puppet to life.
It does all the real work around the weights: chopping your words into tokens, pushing them through the model, choosing each next token, and streaming the answer back to your screen.
Same puppet, different puppeteer, different show. A better harness runs the exact same model faster, squeezes it into less memory, and serves more people at once. The rig matters as much as the doll.
So the model is the puppet, and the harness is what makes it move. You download the weights, but it's the harness that actually performs. That's the harness. Off the Cloud.