No affiliation with mistral, but you don't have to purchase tokens for inference. It runs GPUless and is open source, so you're just paying the normal compute cost.