Glonce

MiniMax-H3 video model ported to MLX, runs on Apple Silicon

modelsThe Glonce editorsSimon Willison🔥65Glonce rating65Impact62Novelty58Relevance92Surprise38Players42Credibility82
MiniMax-H3 video model ported to MLX, runs on Apple Silicon

MiniMax released MiniMax-H3 two days before Simon Willison's post, describing it as "a general-purpose, omni-modal generative system": a model that accepts text, images, audio and video as input and can turn them into video clips up to 15 seconds long, complete with audio. A separate project, PipeNetwork, built a Python package that ports MiniMax-H3 to MLX, Apple's machine-learning framework, so the model can run directly on Apple Silicon hardware instead of requiring a cloud GPU.

Willison tried it himself on his M5 Max MacBook Pro. He cloned the PipeNetwork repository, downloaded both the base MiniMax-H3 model and PipeNetwork's 8-bit MLX-quantized version through huggingface_hub, then ran the included generate.py script with the prompt "a rainbow colored skunk leaps over a mossy log in a supermarket," pointing it at the downloaded model and MLX checkpoint directories. The download pulled in about 115 GB of model files, and generating the single video clip took just under 45 minutes on his machine.

The resulting video itself came out well, which Willison called impressive, but the accompanying audio was, in his words, "weird speech-like garbage." He attributes this to not having supplied any prompt guidance for what the audio should sound like. He notes that a prompting guide with detailed instructions for getting audio generation to work correctly exists, but he had not read it before running this experiment.

Key facts

  • MiniMax released the omni-modal generative model MiniMax-H3 two days before this post, accepting text, image, audio and video input to produce video clips up to 15 seconds long with audio.
  • PipeNetwork built a Python package porting MiniMax-H3 to MLX, letting it run natively on Apple Silicon rather than needing a cloud GPU.
  • Simon Willison ran the MLX port on his M5 Max MacBook Pro, downloading about 115 GB of model files and generating one video clip in just under 45 minutes.
  • The generated video was impressive, but the audio track came out as garbled, speech-like noise because Willison did not supply prompt guidance for the audio, something a dedicated prompting guide covers that he had not read beforehand.

Why it matters

MiniMax-H3 is a new omni-modal generative model that can turn text, image, audio or video input into short video clips with synchronized audio, and within two days of release an independent developer had already made it runnable on ordinary Apple Silicon hardware via MLX rather than requiring cloud infrastructure. Simon Willison's firsthand test shows that a capability class previously tied to cloud GPU access, generating video with audio from a prompt, can now run on a consumer laptop.

Who it affects

Developers and hobbyists running generative AI locally on Apple Silicon Macs are the direct audience, since the MLX port specifically targets that hardware. It is also relevant to anyone tracking MiniMax's product lineup and to developers of MLX tooling, since PipeNetwork's work extends the growing list of models available in Apple's local-inference ecosystem.

How to use it

Running MiniMax-H3 locally requires cloning PipeNetwork's minimax-h3-mlx repository, then using huggingface_hub to download both the base MiniMaxAI/MiniMax-H3 model (specific subfolders only) and the pipenetwork/MiniMax-H3-MLX-8bit quantized version. Generation runs through the package's generate.py script, called with a text prompt and paths to the downloaded checkpoints. In Willison's run this meant about 115 GB of downloads and roughly 45 minutes to produce a single 15-second clip on an M5 Max MacBook Pro. The source gives no pricing or licensing information for either the base model or the MLX port.

How solid is it

This is a firsthand, hands-on account: Willison actually ran the pipeline end to end and produced a real output video, rather than relaying a vendor claim secondhand. That said, it is a single trial on one specific machine, an M5 Max MacBook Pro, with no comparison to other omni-modal or video-generation models and no benchmark or quality metric beyond Willison's own impression that the video looked good.

Risks and caveats

The audio output in Willison's test was unusable, described as weird speech-like garbage, because he did not follow the prompting guidance needed to steer the audio track. That guide reportedly contains substantial detail on getting audio right, but was not consulted for this run, so the demonstrated result understates what the model can do with proper prompting. The source also gives no information on pricing, licensing or commercial availability, and does not say how the MLX port performs on Apple Silicon chips other than the M5 Max.

“a general-purpose, omni-modal generative system”

— MiniMax, quoted by Simon Willison