A video model on a Mac, with a very real download bill

Simon got MiniMax H3 running on an M5 Max MacBook Pro through PipeNetwork’s MLX port. The model accepts text, images, audio and video, and generates clips of up to 15 seconds with sound.

The test asked for a rainbow-coloured skunk jumping over a mossy log inside a supermarket. It required about 115 GB of model downloads and just under 45 minutes of generation.

The video looked impressive. The soundtrack sounded like garbled speech.

That second result came with an unusually useful admission: the prompt had said nothing about audio, and Simon had not read the model’s audio-prompting guidance before running it. A multimodal model may need direction for each part of the output, even when the visual request feels complete.

This was a successful local experiment, with substantial hardware and waiting-time costs. It was not yet the frictionless experience suggested by the phrase “runs on a laptop.”