Skip to content

Free, open-source AI video generator for Mac. Run LTX-2.5 and MiniMax H3 (Hailuo) locally on Apple Silicon: text-to-video and image-to-video with sound, lip-sync, character LoRA training, songs from lyrics (YuE2) and image editing (Qwen-Image-Edit). One-click install via Pinokio. Offline: no cloud, no API key, no NVIDIA GPU.

2.8 KB
Raw Blame

Linux on Apple Silicon - community-maintained, best effort

Contributed by @joshuaswarren (PR #94). macOS through Pinokio is the supported platform; this target is maintained by the community and is not covered by the release gates. Reports are welcome.

Install

From the app root:

bash scripts/linux/install.sh        # engine venv, omarchy-mlx wheel pinned by sha256
bash scripts/linux/install_h3.sh     # optional: the Hailuo H3 engine
./run_panel.sh

The omarchy-mlx wheel is a third-party build; the installers verify its pinned release sha256 before installing it. Models download from the panel, as on a Mac.

Hailuo H3 on Linux

On Linux, H3 can use a complete Q8 DiT pack without the bf16 master. Both config files and all indexed shards must exist; a .built_ok marker alone is not enough. macOS still requires the bf16 master for install discovery.

Measured runs on Omarchy

Linux on Apple Silicon uses MLX with a Vulkan backend instead of Metal. These are separate runs of different engines, not a comparison of machines or Best/Fast controls. Each output was viewed.

Machine Engine and settings Actual output Native job time Peak memory Driver source
14-inch M2 Max MacBook Pro FLUX.2-klein-4B, native queue, browser on another machine 512x512 PNG 58.3 s Not recorded Not captured at run
14-inch M2 Max MacBook Pro H3 native UI, Draft, Fast (3-step), seed 42, 640x384 73 frames, 1280x720 export, 3.042 s duration 1610.79 s Not recorded Not captured; binary hash below
16-inch M1 Max MacBook Pro LTX native queue, 8 steps, seed 42 9 frames, 256x128, 0.375 s duration 181.26 s 31.1G as reported, no swap 69bde15340

The LTX row used an unreleased diagnostics runtime. The published 0.7.32 wheel cannot run its uint8 Take operation, and the installer pin remains unchanged. A requested height of 160 becomes 128 on the 64-pixel canvas grid. The queue also normalized a four-step request to eight steps; this is not a four-step timing.

The H3 run has a viewed 45-second recording of the prompt, Best/Fast controls and finished clip. A Best/Fast timing comparison waits for machine recovery and restoration of the H3 weights.

Driver provenance:

LTX, 16-inch M1 Max MacBook Pro
  source commit  69bde153405cf02d52b0c9810307a0c7d0085d52
  package        omarchy-mlx-vulkan-omarchy 1:26.2.4.git69bde15340-2
  driver sha256  40282811c6e14ef991f72936bb675008cb4d6b3199b1785bc34c7976ab89c425
  kernel         7.1.12-2-12.6-sep-ARCH

H3, 14-inch M2 Max MacBook Pro
  source commit  not captured; this run predates the package above
  driver sha256  a70fb890552f5282ac6ee91b6800d310de9759fe7d18232a48dcbeb31db7c6af
  kernel         7.1.12-2-12.6-sep-ARCH