Final-layout BF16 detector and Object Multiplex tracker weights for running Meta SAM 3.1 with
mlx-cv on Apple Silicon. BF16 is reduced precision, not integer quantization.
1from mlx_cv.models.sam3 import SAM3Processor, SAM3VideoSession
2
3image_model = SAM3Processor.from_pretrained("sam3.1")
4prediction = image_model.predict(image, "person")
5
6video = SAM3VideoSession.from_pretrained("sam3.1")
The strict 1963-tensor BF16 checkpoint loads directly into MLX with no runtime PyTorch conversion. The persisted Metal image gate reached mask IoU 0.999618, maximum box error 0.1626 px, and score error 0.001305 against the official reference. The multiplex decoder mask IoU was 0.99215, with official MPS component checks and real two-frame MLX propagation also passing.
Meta's source Hugging Face repository requires users to accept access terms. This derivative MLX checkpoint is distributed publicly by App Automaton under the bundled SAM License; downloading or using it constitutes acceptance of those terms. Review the complete LICENSE before use.
