Thanks for confirming! Our cold start, excluding model load is 2-4 seconds typically for HF models.

The only time it gets much longer when companies have done a lot with very specific CUDA implementations