h3-metal: MiniMax H3 running natively on Apple Silicon
h3-metal (antirez/h3.c) is an open-source C and Metal inference engine that runs MiniMax H3 on Mac computers. It supports prompt-to-video with audio, first and last frames, and image, video and audio references.
What it does
h3-metal is a native MiniMax H3 engine for Apple Silicon, written in C with Metal shaders, that reads the official Hugging Face snapshot. It's a command-line tool: pass a prompt to render, or run it without one to start an interactive session.
The session keeps the prompt encoding, model and decoder in memory, so trying a new seed doesn't reload everything. Commands like !first and !last set keyframe images.
Results and limits
The README reports end-to-end image+audio and video+audio reference renders in 74.58 and 76.99 seconds on a 128 GB M5 Max, each peaking around 40.1 GB of memory with no swapping. The author describes ongoing performance and memory work on M3 Max and M5 Max.
The README doesn't state a minimum Mac spec.
Worth knowing
The engine is MIT-licensed; the H3 weights use the MiniMax H3 Community License. Users in the EU, UK, US and South Korea need to request a license from MiniMax via the H3 License Request Form (platform.minimax.io/h3-license) before use. FFmpeg and FFprobe must be on your PATH.
Try it yourself
make -j8 && ./h3 --info -d ./MiniMax-H3 ./h3 -d ./MiniMax-H3 --width 512 --height 512 --steps 6 h3> !first [opening.png] h3> !last [ending.png] h3> [The camera moves slowly around the subject.]