H3-metal – Native MiniMax-H3 inference for Apple Silicon

	H3-metal – Native MiniMax-H3 inference for Apple Silicon(github.com)
	438 points by swyx 11 days ago \| 98 comments
	tl;dr: H3-metal is a native Metal implementation of MiniMax-H3 video/audio generation for Apple Silicon (M3/M5 Max), supporting text-to-video/audio, first/last-frame conditioning, and ordered image/video/audio references. It offers extensive speed/quality tradeoffs via layer thinning, step reduction, token reduction, and internal-canvas rescaling, plus an SSD-streaming mode that cuts DiT memory from ~36.5 GiB to ~2 GiB at modest speed cost. M5 hardware gets native BF16 TensorOps and int8 MLP/QKV paths, reducing a 512×512 20-step render from ~36s (BF16) to ~19s (int8) while preserving subject fidelity.
	HN Discussion: ↑Users share existing experience running MiniMax H3 via ComfyUI, noting slow speeds that this native implementation could improve ↑Suggests sparse attention support could provide major speedups beyond current optimizations ↓Notes CUDA/Nvidia hardware still has advantages for diffusion workloads over Apple Silicon ~Concerns about high memory requirements excluding users with less than 128GB RAM •Asks about compatibility and performance on lower-spec Apple Silicon hardware