PipeNetwork/minimax-h3-mlx

Two days after MiniMax dropped its H3 model, someone already had it running on a MacBook. The developer behind this piece describes MiniMax-H3 as an omni-modal generative system, which is a fancy way of saying it takes text, images, audio and video as input and can output up to 15 seconds of video with audio baked in. A Python package called minimax-h3-mlx ports the whole thing to MLX, Apple's framework for running machine learning workloads on its own chips, and Simon Willison decided to try it on his M5 Max MacBook Pro.

The process wasn't exactly casual. Willison cloned the repo, pulled down two separate model downloads through Hugging Face — one for the base model components, another for the 8-bit MLX-quantized version from pipenetwork — and then kicked off generation with a single command line prompt: a rainbow colored skunk leaping over a mossy log in a supermarket. Simple enough as prompts go.

What he didn't mention up front is scale. Getting everything set up meant downloading roughly 115 GB of model files onto his machine, and once that was done, actually generating the clip took just under 45 minutes. That's a real commitment for one short video, and it says something about how much weight these omni-modal systems carry even when quantized down to 8-bit.

The visual result held up. The skunk, the log, the supermarket — it all rendered the way you'd hope. The audio track was another story entirely, coming out as garbled, speech-like noise that didn't match anything happening on screen. Willison admits the failure was on him: he hadn't read the prompting guide beforehand, and that guide apparently contains detailed instructions on how to actually steer the audio generation instead of leaving it to guess.

It's a small but telling example of where these all-in-one generation models currently sit. The video half of MiniMax-H3 seems to just work, even for an offbeat prompt fired off without much thought. The audio half needs real instruction, and skipping that step turns a coherent visual into something with a nonsensical soundtrack bolted on.