MiniMax released its MiniMax-H3 model two days ago, and a community port now allows it to run on Apple Silicon. This Python package enables local execution on M-series chips, accepting text, images, audio, and video to generate clips up to 15 seconds long with sound. Testing on an M5 Max MacBook Pro required downloading approximately 115 GB of files, and generating a video prompt took just under 45 minutes. The resulting clip showed a rainbow coloured skunk leaping over a mossy log inside a supermarket, demonstrating the system’s visual capabilities.
The practical value lies in bringing large omni-modal models to local hardware without relying on cloud APIs. However, the current output requires significant user guidance to avoid audio issues. Without specific instructions on sound design, the generated speech resembles unintelligible garbage rather than coherent dialogue. The model also demands substantial storage space and patience for long generation times on consumer devices.
* Local execution requires downloading roughly 115 GB of model weights.
* Generation of a single clip takes approximately 45 minutes on an M5 Max chip.
* Audio quality degrades without explicit prompt guidance for sound effects.
