Redis Creator Launches h3-metal for Native MiniMax-H3 Inference on Apple Silicon

Salvatore Sanfilippo, the creator of Redis, has released h3-metal, a native inference engine designed to run the MiniMax-H3 multimodal model on Apple Silicon. The project leverages Apple's Metal framework to enable high-performance, local generation of high-resolution video and audio.

The engine provides a native implementation for MiniMax-H3, an open-weights model capable of producing 2K resolution video paired with native stereo audio for durations up to 15 seconds. According to the project's GitHub repository, Sanfilippo is developing the tool in "vertical slices," a methodology that has already resulted in functional prompt-to-video, prompt-to-audio, and frame conditioning capabilities.

The Push for Native Performance

MiniMax-H3 is a general-purpose multimodal model designed to blur the lines between text, image, video, and audio tasks. While other local options exist—such as 8-bit quantizations using the MLX framework—h3-metal distinguishes itself by focusing on a low-level, high-efficiency implementation. By utilizing C and Metal rather than relying on high-level Python wrappers, the project aims to maximize the hardware potential of macOS.

Hardware Optimization and Privacy

This shift toward local inference significantly reduces the industry's reliance on cloud-based APIs, offering creators greater privacy and the ability to iterate on 2K video assets directly on their own machines. To ensure maximum efficiency, the project is currently prioritizing Metal performance and memory optimizations specifically for the M3 Max and M5 Max chips.

What to Watch

As development continues, the focus remains on refining memory management to handle the intensive demands of 2K multimodal generation. Future updates will likely expand the "vertical slices" of functionality, further bridging the gap between professional cloud-grade video generation and local workstation capabilities. This approach signals a broader trend toward bringing high-fidelity generative AI out of the data center and onto the desktop, empowering creators with immediate, private, and high-resolution output without the latency or cost of subscription-based cloud services.