H3cApp

H3cApp for macOS

AI video with sound, made on your Mac.

H3cApp is a macOS app that turns text and images into short videos with sound, using MiniMax H3, a video generation model by MiniMax. Generation runs on your Apple Silicon Mac. It is a ready-to-use app, so there is no Python or PyTorch environment to set up.

Download the latest version How to get started

For Apple Silicon Macs with macOS 15 Sequoia or later. The MiniMax H3 model is a separate download (about 144 GB and up for the official model, or about 35–45 GB as a quantized model (GGUF)) and comes under MiniMax H3's own license.

The H3cApp window in English. The preview shows a generated landscape video of a person playing guitar in a stone plaza at dusk, with Export Video…, Settings Used…, Use as Reference and Show in Finder buttons above it, and the collapsed prompt panel with a summary line and the up-arrow generate button below.
The real H3cApp v0.4.4 window, showing example 3 (the panel shrinks while generating).

Updates

Examples

These videos were generated with H3cApp v0.4.4 (model: MiniMax H3) from the prompts below. All of them have sound; press play to hear it (nothing plays automatically).

Square, Medium (512×512), 5 s, with sound

From text, 20 steps, speed “Standard”, seed 7. Took 5 min 59 s (M5, 24 GB).

A small orange cat plays with a ball of yarn on a wooden floor in soft afternoon light. Static camera, gentle movement. Soft room ambience, no speech, no music.

Landscape, Medium (generated at 672×384, upscaled 2x to 1344×768), 5 s, with sound

From text, 20 steps, speed “Standard”, seed 7. Took 6 min 9 s (M5, 24 GB). The 2x upscale is high-quality resampling after generation.

A street musician plays an acoustic guitar in a quiet stone plaza at dusk, warm lights in the windows behind him. Slow push-in camera. Gentle guitar melody, distant footsteps, no speech.

Square, Small (256×256), 5 s, with sound

Same prompt, seed and settings as the first cat, only the size set to Small. This is the first setting tried in Getting started; details are rougher. Took 1 min 44 s (M5, 24 GB).

The videos are re-encoded for the web. The same prompt and seed can give different results on another Mac or with other settings. The author also has a video on YouTube (external site).

What you can do

Will it run on my Mac?

Check the supported system and free disk space before you download. The AI model in particular is very large and is downloaded separately from the app.

Mac A Mac with Apple Silicon (an M-series chip). Intel Macs are not supported.
macOS macOS 15 Sequoia or later
With an M5-class GPU The fastest compute mode, “Fast (int8 cache)”, and “Fast mode (experimental)” are available.
Other Apple Silicon Runs with “Low memory (SSD streaming)” or “Resident (for plenty of memory)”. Speed and memory needs depend on the Mac and the settings, and are not M5-equivalent.
Memory The author tested on an M5 with 24 GB of memory (512×512 up to 15 s fit in 24 GB). Less memory has not been tested.
App DMG, about 19.0 MB (signed and notarized)
AI model (from text, first/last image) MiniMax H3 FL2VA: about 144.1 GB (134.2 GiB). Needed first.
AI model (reference images, videos, audio) MiniMax H3 Ref2VA: another 144.1 GB or so, about 288.1 GB (268.3 GiB) in total. Add it when you need it.
Other space About 19.3 GB per int8 cache (when used on an M5-class Mac), generated videos, copies of reference files, and working space.
As of October 10, 2026, version 0.4.4. Model sizes are the totals of the Hugging Face file listing. The app labels sizes “GB” but computes them in GiB (1 GiB = 1,073,741,824 bytes), so FL2VA shows as about 134.2.

Mind the model's license. MiniMax H3, the model that generates the videos, comes under its own, somewhat unusual license, the MiniMax H3 Community License, separate from the app's MIT License. It differs from typical open-source licenses: for example, it only applies in certain regions. Before downloading and using the model, read the key points of the license and the full license.

Downloading the model the first time needs an internet connection; generating videos runs on your Mac. See Getting the model for details.

How it goes

  1. Install the app

    Download the DMG from the GitHub releases page and drag H3cApp into Applications.

    Installing

  2. Get the model

    From the window shown at first launch, download the model (about 144 GB). Depending on your connection this can take several hours or more. An interrupted download resumes.

    Getting the model

  3. Make a video

    Make one small 5-second video, play it, and save it.

    Your first video

How long it takes

Measured on the author's Mac. Other settings change both the time and the result.

Video length Speed “Standard” Speed “Fastest” + Fast mode (experimental)
5 s5 min 59 s2 min 38 s
10 s15 min 56 s5 min 39 s
15 s30 min 08 s9 min 25 s
 M5, 24 GB / square 512×512 / from text / 20 steps / compute mode “Fast (int8 cache)” / prompt “A cat playing with a ball of yarn.”, seed 7 / one run per setting (measured October 2–3, 2026). Time from pressing generate to the finished file; the model download is not included. “Standard” and “Fastest” skip different amounts of work, so this does not compare the same output.

Other sizes and speed settings are in the README's measured generation times.

Download

The latest version is on the GitHub releases page. As of October 19, 2026, the latest is v0.6.0 (released October 19, 2026).

Download the latest version Installing

Under Assets on the releases page, choose the .dmg file, such as H3cApp-0.6.0.dmg. “Source code (zip)” and “Source code (tar.gz)” are for developers and are not used for a normal install.

See each release's notes for what changed.