An independent developer has just announced the porting of Meta's SAM (Segment Anything Model) models to the MLX framework, specifically optimized for Apple Silicon. Initial results show a 1.25x increase in inference speed compared to standard methods when running the sam2.1-small model.
Developments
The project is now open-source on GitHub (sam2-mlx) and the model has been shared on Hugging Face. According to the author, the current version uses the fp32 format, but quantized versions will soon be released to reduce size and further increase speed. MLX is Apple's specialized array framework that fully leverages the power of Unified Memory on M-series chips.
Why It Matters
Optimizing computer vision models like SAM on Apple Silicon enables developers in Vietnam to process high-quality images and videos directly on their personal laptops without needing expensive GPU servers. With a 25% performance boost, virtual reality applications or real-time video editing will become significantly smoother on the macOS ecosystem.