Seeing the same on iPhone 18 Pro Max (iPhone19,7), iOS 27.2 and it also affects Apple’s own Vision framework, not only custom Core ML models. We use both. Vision: the first VNDetectFaceLandmarksRequest on a 2316×3088 photo took about 678 s (11 min) in my app. Afterwards it takes 0.1 s, including from a brand-new test app, so it seems like the compiled cache is shared across apps. I tried to fallback to CPU but forcing the request to the CPU via Forcing VNDetectFaceLandmarksRequest onto the CPU with setComputeDevice(.cpu, for:) fails immediately with com.apple.Vision Code=9: “Could not create/add network to inference plan – Espresso exception: ‘I/O error’: Missing weights path landmarksflow-…_quantized.espresso.weights_nonane.” Our own model: loading an enumerated-shape .mlmodelc with computeUnits = .all didn’t return for >10 minutes. The same model with .cpuAndGPU loads in 0.02 s. Is Vision’s face model expected to be covered by the 27.2 improvement? Happy to attach a sysdiagnose.