Antirez/h3.c Project Introduces MiniMax H3 Inference Engine for Apple Silicon

Native Inference for Apple Silicon

The Antirez/h3.c project has introduced a native MiniMax H3 inference engine designed for Apple Silicon. This engine is being developed incrementally, with current work focusing on Metal performance and memory optimization for M3 Max and M5 Max chips.

Key Capabilities

The engine supports prompt-to-video/audio generation, first/last-frame conditioning, and ordered Ref2VA image/video/audio references. These features are functional end-to-end, allowing users to generate and manipulate media based on textual prompts and reference images.

Interactive Session and Conditioning

The h3 binary can be run in an Iris-style interactive session, where users can input prompts to generate numbered videos. The session maintains prompt conditioning, prepared DiT, and video decoder in memory to avoid re-loading for repeated prompts. Users can also apply persistent first/last-frame conditioning or use Ref2VA conditioning images to guide video generation.

Usage and Configuration

To use the engine, users need the Hugging Face snapshot in the specified directory and FFmpeg/FFprobe available on their PATH. The command-line interface provides options for checking model layout, selecting Metal devices, and configuring generation parameters such as width, height, and steps. A balanced preset is available, generating 22 frames at 24 fps with visual feedback during denoising.