antirez releases 'h3.c', a Metal implementation of the MiniMax H3 video generation model

h3.c is an inference engine reimplemented in C and Metal to run MiniMax's video and audio generation model, H3, directly on Apple Silicon. It was released by antirez (Salvatore Sanfilippo), known as the developer of Redis.

Key Features and Optimization Capabilities

• Direct Implementation: By writing Metal kernels directly instead of using MLX-based wrappers, it provides various "handles" that allow users to directly adjust the balance between quality and speed.
• Fine-Grained Control: Inference performance can be optimized through parameters such as the number of denoising passes (--steps), the number of layers (--layers), and token reduction (--token-reduction).
• SSD Streaming: Using the --ssd-streaming option allows model weights to be read directly from the SSD, significantly reducing memory usage (e.g., from 36.5GiB to 2.1GiB on M5 Max). However, execution speed becomes somewhat slower.

⚠️ Important: License and Regional Restrictions

• While the h3.c code itself is free to use, the MiniMax H3 weights required for execution are subject to regional restrictions.
• South Korea, the United States, the European Union, and the United Kingdom are explicitly listed as Excluded Territories. Therefore, to use this model for actual work or services in South Korea, you must obtain separate license permission via Hugging Face.