MiniMax H3 inference engine optimized for Apple Silicon

The MiniMax H3 inference engine has been developed to optimize AI model performance on Mac computers using Apple Silicon. The project focuses on leveraging the capabilities of Apple's CPU and GPU architectures through incremental optimizations for M3 Max and M5 Max chips.

Key technical features include the utilization of Metal 4 and TensorOps for M5 GPUs, alongside native BF16 support for QKV projections and attention mechanisms. These implementations aim to increase computational efficiency, reduce inference time, and improve memory management.

The engine supports various user controls, such as token reduction to accelerate the process and specific parameters like layers and reuse to balance video detail with generation speed. It also provides capabilities for prompt-to-video and prompt-to-audio generation, including first and last-frame conditioning.