MiniMax-H3 Omni-Modal Model Ported to Apple Silicon via MLX
On August 4, 2026, developer Simon Willison introduced a new Python package that integrates MiniMax's MiniMax-H3 model with Apple's MLX framework, ensuring it works seamlessly on Apple Silicon Macs. The MiniMax-H3 model, promoted as a multifaceted generative tool, can create content across text, images, audio, and video formats, producing clips of up to 15 seconds with audio. During a demo on his M5 Max MacBook Pro, he noted that generating a video from a specified prompt required approximately 115 GB of data and took around 45 minutes. He received feedback on the audio quality, which was deemed unsatisfactory.
Key facts
- MiniMax released MiniMax-H3 two days before the post.
- MiniMax-H3 is described as a 'general-purpose, omni-modal generative system'.
- It accepts text, images, audio, and video inputs.
- It can generate up to 15-second video clips with audio.
- Simon Willison created a Python package to port MiniMax-H3 to MLX.
- The model ran on an M5 Max MacBook Pro.
- The model downloaded ~115 GB of model files.
- Video generation took just under 45 minutes.
- The generated video for the prompt 'a rainbow colored skunk leaps over a mossy log in a supermarket' was impressive, but audio was poor due to lack of guidance.
- The prompting guide contains information for better audio generation.
- Post was published on 4th August 2026.
Entities
Institutions
- MiniMax
- MLX
- Apple