halation

Halation is a native macOS application for generating video locally using MiniMax H3 on Apple silicon. It combines a SwiftUI front end with an MLX back end, ensuring that no data leaves the machine: no account, API key, or upload is required. The app is split into two halves: a SwiftUI app that manages the model library, render queue, format validation, H.265 encoding, and video library, and a Python sidecar that drives the PipeNetwork/minimax-h3-mlx port. They communicate over NDJSON on stdout, with the app owning the user interface and the sidecar handling the tensors.

Rendering is computationally intensive, taking hours per clip. The queue is designed to survive quitting and provides a live seconds-per-step ETA. Setup requires about 105 GB of disk space, primarily for the Qwen3-VL-32B text encoder. The app requires an Apple silicon Mac with at least 96 GB of unified memory, macOS 26 or later by default, and ffmpeg. It can be built and run with `make run`, and distributed as a DMG.

The model has specific capabilities and limitations: fixed 24 fps, 768 px short edge resolution, 5–15 second durations, and joint audio generation at 32 kHz stereo. Guidance and negative prompts are not available due to CFG-distilled weights. Output clips are saved as H.264 with AAC audio, accompanied by a .halation.json sidecar for reproducibility. The app is licensed under AGPL-3.0, and the weights are downloaded separately under their publisher's license.