https://ai.meta.com/blog/segment-anything-model-3/ Though "30ms on an H200 GPU". Presumably that's then multiple seconds on Strix or consumer class GPU.