As an FYI the Meta-Llama-3.3-70B-Instruct now supports up to 128k context lengths. Please note as always if TTFT is your main concern working with the shortest context length would always be the fastest. Also remember you do not have to pass in the context length to the model name i.e Meta-Llama-3.3-70B-Instruct-k . Meta-Llama-3.3-70B-Instruct will route properly to the proper model varietal.
-Coby