H3-metal – Native MiniMax-H3 inference for Apple Silicon
Points and comments are a snapshot, not live.
antirez's h3-metal runs MiniMax-H3 video inference natively on Apple Silicon with Metal.
The project is a sequence of working vertical slices: model metadata, Metal block parity, prompt encoding, and prompt-to-video/audio. It supports first/last-frame conditioning and ordered Ref2VA references. Presets range from 4-step aggressive (3.5s on M5 Max) to 50-step reference. Controls include denoising passes, layer count, token reduction, and internal canvas size. Resolution must be multiples of 32, with 512x512 validated. FFmpeg is required for media output. Peak memory footprint is about 40 GB on 128 GB M5 Max.
What commenters are saying
Commenters confirm the model runs well on 64 GB Macs using GGUF quantizations (e.g., Q5_K_M), though at reduced speed (over an hour for a 9-second clip on M5 Pro). antirez's implementation is faster than ComfyUI workflows. One user notes the 5090 RTX does the same in 2 minutes. Two camps: those praising the open-weight release and those noting diffusion models remain compute-bound, favoring Nvidia GPUs. A request for sparse attention support is noted.