Tools

Running MiniMax-H3 via MLX on Apple Silicon: 15‑second multimodal videos with a 115 GB model

A community package ports MiniMax-H3 to MLX so the model can run on Apple Silicon.

Running MiniMax-H3 via MLX on Apple Silicon: 15‑second multimodal videos with a 115 GB model

AI Newsroom · 4 August 2026 · 2 min read

A PipeNetwork Python package, named MiniMax-H3-MLX, ports MiniMax-H3 to MLX so the model can be run on Apple Silicon. MiniMax describes MiniMax-H3 as "a general-purpose, omni-modal generative system" that accepts text, images, audio and video as inputs and can generate video clips up to 15 seconds long with audio included.

How the author tested it

The author ran the setup on an M5 Max MacBook Pro. The workflow used was:

  1. Download the required model files from the Hugging Face hub:
uvx --from huggingface_hub hf download MiniMaxAI/MiniMax-H3 \
  --include 'FL2VA/*' --exclude 'FL2VA/transformer/*'
uvx --from huggingface_hub hf download pipenetwork/MiniMax-H3-MLX-8bit
  1. Run the model with MLX:
uv run --with mlx-vlm \
  --with-requirements requirements.txt python scripts/generate.py \
  "a rainbow colored skunk leaps over a mossy log in a supermarket" \
  -o skunk.mp4 \
  -c ~/.cache/huggingface/hub/models--MiniMaxAI--MiniMax-H3/snapshots/fa9c8ab1eaa21c8ae25e7e40b83b2e6002f340af/FL2VA \
  -t ~/.cache/huggingface/hub/models--pipenetwork--MiniMax-H3-MLX-8bit/snapshots/3ac52081470b0488921c3ec3ba84a39097bf2361

The produced test video depicted a rainbow-colored skunk leaping over a mossy log in a supermarket.

Size and runtime

Running the experiment required downloading roughly 115 GB of model files. Generating the video (up to the 15‑second limit) took the author just under 45 minutes on the M5 Max machine.

Quality and notes

Visually, the output was impressive, but the generated audio sounded like odd speech-like noise. The author attributes the poor audio to the lack of specific audio prompting. A prompting guide for MiniMax-H3 exists and provides detailed advice on how to improve results; the author acknowledged they did not consult it before this run.

Why this matters

Porting MiniMax-H3 to MLX enables local multimodal experiments on Apple Silicon hardware, useful for developers and researchers who prefer on-device inference. At the same time, the model's large download size (~115 GB) and nontrivial generation time highlight the resource requirements for local multimodal generative workflows and the need for careful prompting to obtain coherent audio and other modalities.

Tags

ai, generative-ai, mlx, text-to-video, minimax

Related stories