Running MiniMax-H3 Omni-Modal Model Locally on Apple Silicon via MLX
Simon Willison demonstrated how to run MiniMax-H3, a newly released omni-modal generative model, locally on Apple Silicon using the PipeNetwork MLX port. The setup successfully generated a short video with audio from a text prompt on an M5 Max MacBook Pro. Porting this advanced omni-modal model to Apple's MLX framework enables developers to experiment with local video-and-audio generation without relying on cloud APIs. This advances the capabilities of local AI development on consumer-grade hardware. The local execution required downloading approximately 115 GB of model files, and generating a single video clip took nearly 45 minutes. While the video quality was impressive, the generated audio was distorted due to a lack of specific audio prompting guidance.
## BACKGROUND
MiniMax-H3 is an open-weight, omni-modal generative system capable of processing text, images, audio, and video to generate up to 15-second videos with native stereo audio. MLX is Apple's open-source machine learning framework designed specifically for efficient model training and inference on Apple Silicon.