A Method for Running MiniMax H3-Image on MacBooks with Ultra-Low Memory (~3GB for 1024²)
Hi MiniMax team, thanks for open-sourcing H3-Image!
I've built h3-image-mlx, an open-source MLX implementation for Apple Silicon, supporting text-to-image generation, multi-reference editing, and ultra-high-resolution 4K output.
The goal is to make this massive ~67B parameter model run efficiently on Macs, with extremely low memory usage and almost no speed penalty.
Some results:
- 16GB M5 MacBook Air: 1024² in as little as 34.8s, with ~2.5GB peak process memory
- 48GB M5 Pro: 1024² in 23.61s with just 3.23GB peak process memory
- 4K (4096×4096) on M5 Pro: 317.29s with only 7.18GB peak process memory
- ~85% less peak memory than resident inference, with virtually no speed penalty and pixel-identical outputs at 1024²
The implementation combines staged inference, streamed DiT loading, asynchronous evaluation, weight prefetching, and W8A8 Metal acceleration.
All without additional quantization or quality loss beyond the original INT8 checkpoint.
The repository includes a Python API, CLI, automatic model downloads, detailed benchmarks, and source code.
GitHub: https://github.com/Anionex/h3-image-mlx . I'd love for you to give it a try! Feedback, issues, and contributions are always welcome
Thanks again to the MiniMax team for open-sourcing this amazing model :)
