Also, a GPU instance needs 5s to start. The test depends on how large the model is. So a "very small weak model" can lead much faster than 20s