I'm seeing a lot of options to use the new llama models through an API endpoint but which is the best way? For my use case I would want to first test the different models to see which is best and then ideally be able to use one of them at a large scale for cheap.

I may also need to fine tune some of the models so something that also provides an easy way to fine-tune one of the llama models and then serve that fine-tuned model would be nice. Looking for the most cost-effective options for now but would be interested in understanding features that more costly options would offer.