h3-metal brings native MiniMax-H3 video generation to Apple Silicon
Interactive text-to-video with first/last-frame controls runs at 24fps on M3 Max
h3-metal is a new open-source project delivering native MiniMax-H3 inference for Apple Silicon, built as a sequence of working vertical slices. The current milestone brings prompt-to-video/audio, first/last-frame conditioning, and ordered Ref2VA image/video/audio references to end-to-end functionality. Developer focus is now on incremental H3-specific Metal performance and memory optimization on M3 Max and M5 Max, leveraging Apple's unified memory architecture to keep model weights and activations resident without CPU copies.
The interactive CLI runs an Iris-style session where you type prompts to generate numbered videos. First/last-frame anchors via !first and !last persist throughout the session, and Ref2VA references are exposed as <Picture N> tokens with ordered management via !refs. A balanced preset produces 22 frames at 24fps (~0.92 seconds) with 20 denoising steps, 45 of 50 transformer layers, and reuse extrapolation for speed. For quick iterations, --steps 4-7 uses the low-budget winning schedule. Optional --show previews frames in supported terminals via a resident VAE costing ~10 GiB memory. Built with make and Hugging Face snapshot, it's designed for serious local video experimentation on Apple hardware.
- Native Metal inference for MiniMax-H3, optimized for M3 Max and M5 Max with incremental performance tuning
- Interactive session supports prompt-to-video/audio, persistent first/last-frame anchors, and ordered Ref2VA image references as <Picture N> tokens
- Configurable denoising (--steps), layer skipping (--layers), and reuse extrapolation (--reuse) to balance speed and quality
Why It Matters
Enables professional-grade local video generation on Apple Silicon, cutting cloud dependency and enabling fast iterative workflows with unified memory.