Sources
MiniMax-H3 Now Runs on Apple Silicon via MLX — 115GB of Weights and 45 Minutes on an M5 Max for 15 Seconds of Video
Simon Willison tried PipeNetwork/minimax-h3-mlx on August 4, 2026, an MLX port of MiniMax's self-described "general-purpose, omni-modal generative system" that takes text, image, audio and video in and emits video clips up to 15 seconds with audio. The honest numbers: ~115GB of model files to download and just under 45 minutes to generate one clip on an M5 Max MacBook Pro. Willison called the video output impressive but the audio "weird speech-like garbage" — because he supplied no audio guidance and hadn't read the prompting guide first. This is the current floor for local omni-modal video generation on a laptop: it works, it is not yet a workflow.
↳ Follow the thread