10 tokens / second is what you get running llama-30b entirely on the GPU. A 65b model will be slower than that since there's more compute involved.