MiniMax H3 Finally Runs Natively on Mac, Thanks to One Developer

MiniMax H3, the open-weights AI video and audio generation model, had no official way to run on Mac hardware until this week. A developer outside the company built one anyway.

Salvatore Sanfilippo, known for creating the Redis database, released h3.c on GitHub: a native inference engine that runs MiniMax H3 directly on Apple Silicon through Metal. MiniMax never shipped an official MPS, Metal, or MLX recipe for the model, and no verified Apple Silicon benchmark existed before this release.

What h3.c Actually Does

H3 is a dense, 33-billion-parameter diffusion transformer, not a small model to run casually on a laptop. Sanfilippo built h3.c as a sequence of working vertical slices rather than one release: host and model metadata first, then Metal block parity, prompt encoding, and prompt-to-video/audio generation. Prompt-to-video/audio, first/last-frame conditioning, and ordered image, video, and audio references already work end to end. Current development is focused on Metal-specific performance and memory optimization on M3 Max and M5 Max chips.

The project also incorporates code from Liu Liu, developer of the Draw Things app, with Sanfilippo inviting him to pull any parts back for his own project if useful.

Competitive Context

Open weights are supposed to mean anyone can bring a model to new hardware, but that promise usually depends on someone actually doing the work. MiniMax open-sourced H3's weights, and the community has been racing to make it usable outside high-end Nvidia setups since.

Other efforts include 4-bit quantized versions runnable with as little as 8GB of VRAM through DiffSynth-Studio, and an MLX port that trims 13 billion parameters at inference through AdaLN precomputation. Sanfilippo's engine is the first with a working Metal-native path built specifically for Apple Silicon rather than adapted from a CUDA-first codebase.

The Signal in the Noise

This is what open weights are supposed to produce: a capable outsider fixing a real gap instead of waiting for the vendor to prioritize it. MiniMax gets a functioning Mac path it didn't have to build, and Mac-based filmmakers get access to a serious open video model without needing a dedicated GPU rig.

Does community-built infrastructure like this change how much you weigh "open weights" as a factor when picking which AI tools to invest time in?

Specs & Pricing

• Model: MiniMax H3, 33B-parameter dense diffusion transformer, 50 layers
• Engine: h3.c, Metal-native inference for Apple Silicon
• Working features: prompt-to-video/audio, first/last-frame conditioning, ordered Ref2VA references
• Optimization target: M3 Max and M5 Max
• Requirements: Hugging Face MiniMax-H3 snapshot, FFmpeg and FFprobe on PATH
• Cost: free, open source

Resources & Reads

• h3.c on GitHub
• MiniMax-H3 official repository