I've read about yesterday someone running LLAMA in a single GPU. Maybe if you optimise the model enough, you can give it to them as a box.