antirez's h3.c: Running MiniMax-H3 Video Generation Natively on Apple Silicon

Antirez/h3.c: MiniMax H3 inference engine for Mac computers

Antirez's h3.c is a native MiniMax-H3 inference engine for Apple Silicon, implementing prompt-to-video generation with Metal performance optimizations. It supports interactive sessions, first/last-frame conditioning, and Ref2VA references. The project offers a range of speed/quality presets, from a fast 4-step mode that generates a 22-frame video in about 3.5 seconds on M5 Max, to a close-to-reference 50-step mode. It also includes token reduction, internal canvas scaling, and automatic RoPE adaptation for 256x256 output.

The four-pass denoise took about 3.5 seconds on M5 Max, versus 26.4 seconds for the reference.

More from this day

2026-08-11