AI Video Offline on a Mac (Apple Silicon): What Runs, How Long It Takes, What It Needs
You can make AI video on your Mac’s GPU for free. It is slow, short and memory-hungry.
No credits, no watermark from a hosted service and nothing leaves your machine. That is the appeal of running a video model on an Apple silicon Mac. The cost is time, disk space and memory. This guide shows which models run, which Mac memory tier fits each, what people have actually measured, and one trap in the official LTX app.
The short answer
- Two model families run on a Mac: Wan 2.2 and LTX-2.3. You run them through community tools, mainly the Draw Things app or the open-source mlx-video package.
- Memory decides everything. The full Wan 2.2 5B model peaked at 56GB in one test, while shrunk versions of the models were reported on 24GB Macs.
- Expect short clips. The reported runs make roughly 2 to 5 seconds of video and take from a couple of minutes to over an hour each.
- The trap: LTX’s own Desktop app on a Mac does not run locally. LTX’s help pages say Mac generation uses their paid API.
Is this the right guide for you?
This guide covers one use case: generating clips privately on your own Mac, with no credits. If you want a free clip today from a website, read our guide on free hosted AI video tools instead. If you want finished animated videos such as titles, explainers and reels, you don’t need AI generation at all, and our HTML-to-MP4 guide is the better route.
What “offline video” means here
A video model works on your Mac’s GPU, which shares the computer’s main memory (Apple calls it unified memory). So the amount of memory your Mac has sets how big a model can run. Model files are also huge, so a first run means a large download. After that, Draw Things says local creation works offline, and mlx-video runs from files on your disk.
The ways to run it
| Route | What it is | What the official page says | Best for |
|---|---|---|---|
| Draw Things | A Mac, iPhone and iPad app from Draw Things, Inc. | Free with in-app purchases. Lists Wan 2.2 and LTX-2.3 among its models. Says local creation has no ads or token limits, with optional cloud compute | Beginners. No command line. We found no published video speeds or memory needs from the developer, so test it |
| mlx-video | An open-source package (MIT licence) built on Apple’s MLX framework | Supports LTX-2 and LTX-2.3, and Wan 2.1 and 2.2. Needs macOS on Apple silicon, Python 3.11 or newer and MLX 0.22 or newer | People comfortable with Terminal. Some of the published Mac timings use it |
| ComfyUI | A node-based image and video tool | No Mac guarantee. Reports on GitHub describe a float8 crash on an M2 Mac and corrupted output on an M4 Max | Experts only. Not where to start |
| LTX Desktop | LTX’s own video editor app | On macOS it uses API mode only. Local GPU generation is not supported on Mac, an API key is required and generation uses paid credits | Not offline. Skip it if you want no credits |
What people have measured on a Mac
No chip maker or model maker publishes Mac timings for these models, so the numbers below come from community write-ups and project pages. We have not re-run them. They are not a chip comparison, because each row uses a different model, resolution and setting. At 24 frames a second, 81 frames is about 3 seconds of video, and 121 frames is about 5 seconds.
| Mac | Model and route | Output | Time | Peak memory |
|---|---|---|---|---|
| M3 Max, 64GB | Wan 2.2 5B, full model, mlx-video, text to video | 1280×704, 41 frames | 17 min 43 s | 56.4GB |
| M3 Max, 64GB | Wan 2.2 5B, full model, mlx-video, image to video | 1280×704, 121 frames | 96 min | 57GB |
| M4 Max, 36GB | FastMetal 5B (a 3-step shrunk and distilled Wan 5B) | 1280×704, 81 frames | 151 s | 9.34GiB |
| M5 MacBook Air, 24GB | FastMetal 5B | 1280×704, 81 frames | 200 s | 9.54GiB |
| M5 Pro, 24GB | LTX-2.3, int8, ltx-2-mlx-swift | 704×512, 121 frames | 150 s | Not stated |
| M5 Max, 128GB | Wan 2.2 A14B, 4-bit pack, 10 steps | 512×320, 81 frames | 10 min 25 s | 16.37GiB |
Which Mac memory tier fits
None of the model makers publishes a Mac memory requirement, so this is our reading of the reports above. Treat it as a starting point.
- 16GB.Only small, shrunk builds are listed for this tier, and we found no timing from a 16GB Mac. Expect to be limited.
- 24GB.Published runs exist: a 24GB MacBook Air with M5 ran the shrunk Wan 5B, and a 24GB M5 Pro ran LTX-2.3 at int8. These are the lighter, faster-running versions.
- 32GB.One report shows LTX-2.3 running on an M2 Pro with 32GB, with block streaming and 33 frames at 480×704. The author of the full Wan 5B test wrote that 32GB is not enough for that model.
- 64GB.The full Wan 2.2 5B model ran here, slowly, and peaked at 56.4GB.
- 128GB.The one Mac timing for the larger Wan 2.2 A14B model came from this tier, and only at low resolution.
The LTX Desktop trap
LTX makes an editor app called LTX Desktop that runs LTX-2.3. On Windows and Linux with a supported NVIDIA GPU it generates locally. On a Mac, LTX’s help pages say it does not: generation uses LTX’s cloud API, needs an API key and uses paid credits. So if a post tells you LTX Desktop gives you free local video on a MacBook, check the system requirements page first. For local Mac generation, use Draw Things or mlx-video instead.
What it takes before you start
- Apple silicon. mlx-video lists macOS on Apple silicon as a requirement, so an Intel MacBook is out for that route.
- Disk space. ModelFit’s reading of the Hugging Face file listings shows a Wan 2.2 5B set of about 18GB in ComfyUI’s format and about 34GB from the official repo. A complete LTX-2.3 set ranges from about 27GB (a shrunk GGUF build) to about 70GB (the official full-precision files). The larger Wan A14B official repo is about 126GB.
- Patience. Even the best reported full-model run took 17 minutes for under 2 seconds of video.
A sensible first test
- Install Draw Things from the Mac App Store. It is free, with in-app purchases.
- Pick the smallest video model it offers and a low resolution. Don’t start with the biggest model.
- Make one short clip, and watch memory use in Activity Monitor while it runs.
- Note the time. If one clip takes longer than you would wait for a finished video, a hosted tool or the HTML-to-MP4 route is a better fit.
Licences
- Wan 2.2: the Wan repository says its models are Apache-2.0 licensed, which allows commercial use under its terms.
- LTX-2.3: LTX’s licence page says commercial use is free under the community licence for companies earning under $10M in annual revenue, and needs a commercial licence above that. The revenue is for the whole company.
- Shrunk or derived builds (quantised, distilled or converted) carry their own licence cards. Read the card for the exact build you download.
Questions we get asked
Can a 16GB MacBook make AI video offline?
Only with small, shrunk models, and we found no timing from a 16GB Mac. Reported runs come from 24GB and above. Try a short test before you plan a project around it.
Can an Intel MacBook do this?
Not through mlx-video, which needs Apple silicon. Intel Macs are better used for editing and for cloud or code-based tools.
Is it really free?
The software and models are free under their licences. You pay in time, disk space and electricity. Draw Things also sells optional cloud compute, and LTX Desktop on a Mac uses paid API credits.
Does it work without internet?
After you download the models. Draw Things describes local creation as offline. The downloads themselves are large.
Is it faster than hosted tools?
No. Hosted tools run on data-centre GPUs. A Mac’s advantage is privacy and having no credits, not speed.
Choosing a Mac for AI video?
Tell Aparna what you plan to make and your budget. She’ll say honestly whether a Mac with more memory, or a cloud tool, is the better fit, and which MacBooks in stock are Apple silicon.
Tools, models and figures checked 28–30 October 2026 and change often. Timings are community-reported and were not re-run by us. Model and app availability can differ by region.
