antirez ships h3.c: MiniMax H3 inference on Apple Silicon

A one-developer project called h3.c has landed on GitHub, and it is the sort of thing that quietly rearranges assumptions about what a Mac can do at inference time. The repo is a native Metal implementation of MiniMax's H3 multimodal generation model, built by antirez, targeting Apple Silicon directly rather than sitting on top of a Python framework.

The numbers are the interesting bit. On an M5 Max, the fast preset renders 512×512 in about 16.69 seconds, and the aggressive four-step path finishes in roughly 3.5 seconds. The transformer checkpoint pulled from Hugging Face is about 33 GB, and the README puts peak physical memory at approximately 40 GB during end-to-end generation. Capabilities include text-to-video, text-to-audio, image conditioning, and video continuation, with output capped at 768×1344 pixels at 24 fps and valid frame counts following a 5 + 17n rule.

The optimization work under the hood is what makes the repo technically interesting. h3.c leans on native TensorOps for BF16 matrix operations under Metal 4 on M5, adds an int8 quantization path (19.32 seconds versus 25.80 with BF16), and uses activation buffer aliasing to remove 61.25 MB at the 512-class geometry. There is also token reduction via pair-pooling after block 3 and a streamed Qwen text encoder with prefetching. Three researchers in our AI Weekly Who's Who directory circulated the source link, which lines up with how much engineering is on display.

A few things the README explicitly does not promise. Numerical pixel identity with the reference MLX implementation is not expected, because the random-number and execution engines differ. The tested hardware footprint is narrow, M3 and M5 Max only, and it is a solo project shipped under MIT license. If you do not already own a high-end Apple Silicon Mac with plenty of unified memory, none of this helps today.

For teams building on-device creative tooling on Apple hardware, though, this is a useful data point: a real multimodal generator running locally at seconds-per-clip, from a solo shipper, with no cloud round trip in the loop.