PipeNetwork/minimax-h3-mlx

MiniMax dropped a new model called MiniMax-H3 just two days ago, billing it as an "omni-modal generative system" that can chew through text, images, audio and video and spit out up to 15 seconds of video complete with synced audio. That's a fairly ambitious pitch. What makes this particular release interesting isn't the model itself so much as a quick community port: a Python package called PipeNetwork/minimax-h3-mlx that gets it running via MLX, Apple's machine learning framework built for its own chips.

Simon Willison tested it on his M5 Max MacBook Pro, and the process, while doable, is not for the faint of hardware. Getting the model onto disk required pulling down roughly 115 GB across two separate Hugging Face repositories, one for the base FL2VA components and another for an 8-bit quantized MLX version from PipeNetwork. After that, a single command-line call to a generate.py script handled the actual video creation.

He fed it the prompt "a rainbow colored skunk leaps over a mossy log in a supermarket," which is exactly the kind of absurd test case this stuff deserves. The generation itself took just under 45 minutes on his machine, and the resulting clip looked genuinely good on the visual side. The audio, though, came out as what he described as speech-like garbage, essentially the model guessing at sound without any direction.

That failure wasn't really the model's fault. MiniMax ships a prompting guide with detailed instructions on how to steer the audio generation, and Willison admits he skipped it before running his test. It's a familiar story with these multimodal systems: the raw capability is there, but getting clean, coherent output means learning the prompt syntax first, not just typing a sentence and hoping for magic.

Still, the fact that a five-day-old omni-modal video model already has a working MLX port speaks to how fast this ecosystem moves. A year ago running something like this locally on a laptop, even a maxed-out one, would have sounded far-fetched. Now it's a 115 GB download and a single terminal command away, garbled skunk audio and all.