MLX port released for MiniMax-H3, enabling text-to-video generation on Apple Silicon

MiniMax released MiniMax-H3, an omni-modal generative model that accepts text, images, audio, and video as inputs and can produce up to 15-second video clips with audio. The model is positioned as a general-purpose system capable of multimodal generation. A community-maintained Python package, PipeNetwork/minimax-h3-mlx, ports MiniMax-H3 to MLX, Apple’s machine learning framework optimized for Apple Silicon. This enables users to run the model locally on M-series Macs. The author demonstrated inference on an M5 Max MacBook Pro, executing the model via a provided Python script. Running the model required downloading approximately 115 GB of model files. Generating a 15-second video clip took just under 45 minutes on local hardware. The resulting video was visually coherent, but the audio was described as unintelligible speech-like noise due to the absence of prompt guidance for audio generation. MiniMax provides a prompting guide to address audio quality in future runs. The port exemplifies the growing ecosystem of community-driven optimizations that make frontier multimodal models accessible on consumer devices, reducing dependence on cloud-based inference services.