antirez's h3.c brings native MiniMax-H3 video generation to Apple Silicon

h3.c is a from-scratch C and Metal implementation that runs the MiniMax-H3 generative model directly on Apple Silicon, targeting M3 Max and M5 Max machines. Beyond a single generation command, the binary offers an interactive, llama.cpp-style session that keeps prompt conditioning, the prepared diffusion transformer, and the video decoder resident in memory, so re-rolling a prompt with a new seed skips reloading and re-encoding. It handles prompt-to-video and audio, persistent first/last-frame anchors, and ordered Ref2VA image references, with FFmpeg as the only external dependency.

The real focus is a tunable set of quality-versus-speed controls, each benchmarked against a slow reference render using full-video SSIM. Operators can cut the number of denoising passes, extrapolate skipped velocity steps (—reuse), thin the transformer from 50 to fewer blocks, apply token reduction, render internally at a lower resolution before upscaling, or enable an int8 fast path on M5. The gains are steep: a four-pass fox render finished in roughly 3.5 seconds on an M5 Max against 26.4 seconds for a 29-pass reference, holding about 0.556 SSIM. The documentation is candid about the tradeoffs, listing combinations that produce ghosted limbs, color ringing, or drifting composition.

The project’s significance is less the model than the delivery: a dependency-light native pipeline that puts diffusion-based video generation on consumer Mac hardware without the usual Python and MLX stack, developed as a sequence of verifiable vertical slices. It reflects a broader push toward efficient, local, on-device generative inference, with performance work still ongoing and results explicitly sensitive to thermal throttling.