AI Signal 497
Antirez/h3.c: MiniMax H3 inference engine for Mac computers
Antirez released h3.c, a native Metal inference engine for the MiniMax-H3 multimodal model that runs on Apple Silicon, enabling prompt-to-video generation with interactive sessions and configurable quality-speed tradeoffs.
Running a large multimodal model locally on Apple Silicon removes the dependency on cloud inference for video generation, giving developers direct control over latency, cost, and data privacy. The project demonstrates that meaningful speedups are possible through partial denoising passes and layer reduction without catastrophic quality loss.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
The engine implements MiniMax-H3 inference natively on Apple Silicon using Metal, with current optimization work targeting M3 Max and M5 Max hardware.
Interactive sessions keep prompt conditioning, DiT, and decoder weights resident in memory so repeated prompts avoid reloading and re-encoding.
A four-pass denoising configuration achieved roughly 0.55 SSIM against a 29-pass reference while completing in approximately 3.5 seconds on M5 Max versus 26.4 seconds for the full run.
THE CLUSTER