On-device OCR with dots.ocr via CoreML and MLX on Apple Neural Engine
AI Impact Summary
Proposes running the 3B-dot OCR model, dots.ocr from RedNote, on-device using CoreML and MLX to leverage Apple Neural Engine for power-efficient inference. The approach details PyTorch-to-CoreML conversion steps and targeted fixes (dtype alignment, masking removal, static shapes) to achieve a GPU/Neural Engine path, highlighting migration friction for developers. If adopted, this enables offline OCR with reduced latency and data exposure risk, but success depends on reliable on-device conversion and cross-device compatibility.
Affected Systems
- Date
- Date not specified
- Change type
- capability
- Severity
- info