Community MLX Port Brings MiniMax-H3 to Apple Silicon
PipeNetwork released an MLX port of MiniMax-H3 for Apple Silicon, translating the 33-billion-parameter video-and-audio diffusion pipeline into a native MLX implementation. The repository documents structural memory reductions and quantized checkpoints, while an independent test by Simon Willison still required roughly 115 GB of model files and about 45 minutes to generate a short clip on an M5 Max MacBook Pro.
PipeNetwork released minimax-h3-mlx, a community MLX implementation of MiniMax-H3 for Apple Silicon. The project ports MiniMax's joint video-and-audio diffusion pipeline rather than wrapping it through mlx_lm: the repository describes a 33-billion-parameter diffusion transformer conditioned by a frozen Qwen3-VL-32B encoder, with separate video and audio decoders.
What the port changes
The repository says the two upstream MiniMax-H3 task checkpoints contain the same underlying weights, allowing one conversion to cover both. It also documents two structural reductions for local inference: precomputing the adaptive-layer-normalization projections means 13 billion parameters do not need to remain resident, and the port loads only the first 50 of the encoder's 64 decoder layers because later layers are not used by the pipeline. The maintainers estimate those changes reduce the resident pipeline from about 144 GB to roughly 102 GB before quantization.
Published checkpoints include faithful mixed-precision and quantized variants. The project reports that the full text-to-video-and-audio path has run end to end, while the keyframe-conditioned path is implemented but has not yet received the same end-to-end validation. That distinction matters for developers evaluating the repository: it is usable experimental software, not a claim that every upstream workflow is fully reproduced.
What an independent run showed
Simon Willison reported cloning the repository and running it on an M5 Max MacBook Pro. His test downloaded about 115 GB of model files and took just under 45 minutes to produce a short video. He found the video visually convincing, while the generated audio was poor because his prompt did not include audio guidance.
MiniMax's upstream model card describes H3 as an omni-modal system that accepts text, images, video, and audio and can generate video with native stereo audio for clips up to 15 seconds. The official workflow can reach 2K output through a separate regeneration stage, but the MLX repository cautions that dense attention makes that flagship resolution impractical for local runs.
For Apple Silicon developers, the release is most useful as a transparent reference implementation and an early path for local experimentation. Its real constraint is compute time rather than a simple install-and-run experience, and users should treat the repository's performance and quality measurements as hardware- and configuration-specific.
Key Points
- 1The community port implements MiniMax-H3's joint video-and-audio diffusion pipeline directly in MLX for Apple Silicon.
- 2Repository optimizations reduce resident memory before quantization, but independent testing still required roughly 115 GB of model files and a long generation time.
- 3Text-to-video-and-audio generation has been run end to end; the keyframe-conditioned path remains less fully validated.
Scoring Rationale
The port gives Apple Silicon developers a transparent path to experiment with a large synchronized video-and-audio model, but its storage and runtime demands limit near-term accessibility.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
