Salvatore Sanfilippo published h3.c, an open-source engine written in C that runs MiniMax's H3 video-and-audio generation model directly on Apple Silicon Macs. It generates clips from text prompts, supports first and last frame anchors plus ordered image, video and audio references, and offers presets that trade quality for speed. It requires a local copy of the model weights and FFmpeg, and work on performance and memory use is still ongoing.
What changed
Running MiniMax-H3 video generation on Apple hardware previously relied on the released MLX-based workflow rather than a native C and Metal engine.
What it unlocks
Generating short video with synchronised audio from a text prompt on an Apple Silicon Mac, including first and last frame anchors and ordered image, video and audio references, with a memory-saving mode that streams model blocks from SSD.
- four-pass denoise about 3.5 seconds on M5 Max versus 26.4 seconds for the 29-pass reference
- DiT storage falls from about 36.5 GiB to 2.0 GiB with SSD streaming
- maximum canvas 768 x 1344 pixels; output at 24 fps
What you need to act on it
- Apple Silicon Mac
- local copy of the MiniMax-H3 model files from Hugging Face
- FFmpeg and FFprobe installed
- building the tool from source
- github.com2026-08-11