Didn't see where you got those numbers, but surely that's just a problem of throwing more compute at it? From the blog post:

> This excellent performance comes with fast inference — SAM 3 runs in 30 milliseconds for a single image with more than 100 detected objects on an H200 GPU.