Developer Antirez releases native MiniMax H3 inference engine for Apple Silicon
Salvatore Sanfilippo’s new engine leverages Metal and TensorOps to optimise video generation performance, offering granular control over memory management and thermal throttling for local AI workloads.
Developer Salvatore Sanfilippo, known in the technology community as Antirez, has released h3.c, an open-source C project that provides native inference capabilities for the MiniMax H3 model on Mac computers equipped with Apple Silicon. Hosted on GitHub, the engine is designed to optimise performance specifically on M3 and M5 Max hardware, utilising Apple’s Metal framework and TensorOps to handle complex video generation tasks.
The project supports a range of multi-modal features, including prompt-to-video generation, first and last-frame conditioning, and the integration of image, video, and audio references. Users can interact with the engine through an interactive command-line interface that retains prompt conditioning and model state in memory, allowing for rapid iteration without reloading models when changing parameters such as random seeds.
Performance benchmarks on M5 Max hardware indicate significant speed improvements, with some configurations reducing denoising time from approximately 39 seconds to under 20 seconds. The engine employs specific memory management strategies, such as zero-copy weight mapping on M5 hardware and activation buffer reuse, to mitigate peak memory usage and reduce the risk of thermal throttling during intensive inference runs.
Technical optimisations include token reduction, layer thinning, and core reuse to balance processing speed with output quality. The system also handles audio references between two and 15 seconds, synchronising them with video outputs using native BigVGAN and AudioVAE decoders. While the engine is robust, users are advised that combining certain aggressive tuning settings may result in visual artifacts, such as colour ringing or ghosted limbs.
The release provides extensive command-line options for tuning resolution, denoising steps, and frame rates, catering to developers and researchers seeking fine-grained control over local generative AI workflows. By leveraging the specific architecture of Apple Silicon, h3.c aims to make high-fidelity video generation more accessible on consumer-grade Mac hardware without relying on external cloud services.
