The previous packages computed in FP32, which the Neural Engine cannot run: Core ML placed them on the GPU (ALL) or the CPU (CPU_AND_NE). Same paths and file names; recipe: source/build-camstack-models.py (half=True).
fix: reduced YOLO26 CoreML replaced in place; drop the -static copies
The static (batch 1) YOLO26 @320/@256 CoreML packages already live at their original paths. The -static copies are removed, and source/ carries the recipe that exports them static under the original names.
fix: YOLO26 @320/@256 CoreML re-exported with a static batch
The batch-flexible (1..16) YOLO26 @320/@256 CoreML packages abort the loading process (SIGABRT in Core ML's Espresso gather_nd shape pass, macOS 26.2, every compute unit, coremltools 8.3 and 9.0). Re-exported static, batch 1, FP16 (Ultralytics 8.4.92, torch 2.5.1, coremltools 8.3.0) with scripts/build-camstack-models.py. Published as <stem>-static.mlpackage and over the old <stem>.mlpackage path for older catalogs.
8 days ago
Squash history: mirror holds only the models CamStack offers (2026-10-06)
