I got MiniMax H3-Image running on a MacBook air with just ~3GB memory and almost no speed loss
retained text
Hey everyone! I've been working on bringing MiniMax's new \~67B H3-Image model to Apple Silicon with MLX. And yes, **it actually runs on a 16GB MacBook Air!** A community contributor tested it on an M5 MacBook Air (16GB): * 1024×1024 in as little as 34.8s * \~2.5GB peak process memory * Original 6-step model, no extra quantization The key is keeping memory usage low without sacrificing speed or image quality. I combined staged inference, on-demand DiT weight streaming, asynchronous evaluation, weight prefetching, and W8A8 Metal acceleration. **On my M5 Pro (48GB), at 1024×1024:** * Resident: 23.63s / 21.18GB * Streaming: 32.33s / 2.53GB * **Streaming + prefetch + async: 23.61s / 3.23GB** That's **\~85% less peak process memory, with virtually no speed penalty!** All three modes produce pixel-identical images. No additional quantization or quality loss beyond the original INT8 checkpoint. It also supports native 4K generation and multi-reference image editing. I've open-sourced the MLX implementation, with a Python API, CLI, automatic model downloads, and reproducible benchmarks. **GitHub:** [https://github.com/Anionex/h3-image-mlx](https://github.com/Anionex/h3-image-mlx) Would love to hear your thoughts, especially if you have a lower-memory Mac and want to give it a try! https://preview.redd.it/ccvq8bqz4b1i1.png?width=2424&format=png&auto=webp&s=664f76837da752f50f5a6ece646a6ac4e7e6b736
document fetched 2026-11-14 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xuyirz/i_got_minimax_h3image_running_on_a_macbook_air