Worked on the mlx-audio repository to deliver a feature that streamlined the model loading process for FireRedASR2. The approach involved removing the previous conversion step, allowing the system to load pre-converted model weights directly from Hugging Face. This change simplified initialization, reduced maintenance overhead, and minimized risks associated with conversion errors. The work leveraged Python for implementation, with a focus on audio processing and machine learning principles. Unit testing was used to ensure reliability of the new loading flow. The update resulted in a more efficient and maintainable workflow for deploying FireRedASR2 models within the mlx-audio project.