An open source AI video generator is a model you download and run yourself. No credits, no monthly plan, no company that can retire it. The price is hardware: these models want a graphics card that costs more than a year of most subscriptions.
This page compares the five families worth knowing going into 2027. The licence and hardware lines come from each project's own repository, read on November 21, 2026.
Which open source AI video generators are worth running?
| Model | Licence, per its repository | Hardware its page lists |
|---|---|---|
| Wan2.2 | Apache 2.0 | 5B model: at least 24GB of GPU memory. 14B models: at least 80GB |
| HunyuanVideo-1.5 | See the repository | Minimum 14GB of GPU memory, with model offloading |
| HunyuanVideo (original) | See the repository | Minimum 60GB for 720p, 45GB for 544 by 960 |
| LTX-2 | Code under Apache 2.0, models under the LTX community licences | See the repository. It has options for limited GPU memory |
| CogVideoX | Code under Apache 2.0. The 2B model is Apache 2.0, the 5B model has its own licence | 2B model: from 4GB with memory optimisations |
| Mochi 1 | Apache 2.0 | About 60GB on a single GPU. Less than 20GB through ComfyUI |
Wan2.2
Wan2.2 is the default answer. The models are licensed under Apache 2.0, and the project says it claims no rights over what you generate. The TI2V-5B model does text to video and image to video at 720p and 24 frames a second, and its page says it runs on a GPU with at least 24GB of memory, naming the RTX 4090. The larger 14B models ask for at least 80GB, with options that reduce memory use at the cost of speed.
HunyuanVideo and HunyuanVideo-1.5
Tencent's first HunyuanVideo was heavy: a minimum of 60GB of GPU memory for 720p and 80GB recommended. HunyuanVideo-1.5 is the one to look at now. Its repository lists a minimum of 14GB of GPU memory with model offloading enabled, which puts it within reach of far more graphics cards.
LTX-2
LTX-2 from Lightricks is the family that replaced the older LTX-Video. One detail matters for commercial work: the code in the repository is open source under Apache 2.0, but the models are not part of that. They are released under their own LTX community licences. Read the model licence before you sell anything made with it.
CogVideoX
CogVideoX is the lightest entry. Its page lists the 2B model from 4GB of GPU memory when run with the diffusers optimisations, and the 2B model's licence was changed to Apache 2.0. The 5B model is under a separate CogVideoX licence.
Mochi 1
Genmo released Mochi 1 under what it calls a permissive Apache 2.0 licence. On a single GPU the repository says it needs about 60GB of memory, and that ComfyUI can optimise it to run on less than 20GB.
What do you run them in?
Most people use ComfyUI, an open source app that runs these models as a graph of nodes. Its repository says it is available on Windows, Linux and macOS, with a desktop installer for Windows and macOS, and that it supports NVIDIA cards, AMD, Intel Arc and Apple silicon. It also has memory management that offloads parts of a model, which is how large models fit on smaller cards.
Expect setup time. You download gigabytes of weights, pick a workflow, and tune settings before the first good clip.
Is open source really free?
In money per clip, yes. In total cost, count three things:
- The graphics card. 24GB of memory is the comfortable entry point for Wan2.2's 5B model.
- Your time. Installing, updating and fixing a local setup is part of the deal.
- The licence. "Open source" on a headline does not always cover the model weights. Wan2.2 and Mochi 1 state Apache 2.0. LTX-2's models and CogVideoX's 5B model have their own terms.
If you do not own the hardware, a hosted plan is cheaper for occasional use. The best free AI video generator lists what the hosted free plans include, and Sora alternatives covers the paid ones.
What if you want local video without a big graphics card?
All the models above generate footage with diffusion, and that is what needs the memory. A different kind of local tool does not.
Easy Reels is a desktop app for Mac and Windows that makes finished reels on your computer: drawn looks such as pixel art, ASCII, blueprint and low poly, and narrated Motion reels with a voice, captions, music and sound effects. It writes and renders the video as animation, so it runs on a Mac with Apple silicon or a Windows 10 or 11 PC without a 24GB card.
Be clear about the differences. Easy Reels is not open source: it is a paid app, $39 once, with a 7-day refund. And it does not generate photorealistic footage, so Wan2.2 stays the pick for a film shot. Easy Reels covers the script, voice, captions and edit of an explainer. It can run on a free local model for narrated reels and for editing your own footage. The drawn looks need Claude Code, Codex or Grok. The local AI video generator page has the detail, and Can Claude make videos explains the approach.
Questions
What is the best open source AI video generator?
Wan2.2 for most people: Apache 2.0 licence, and a 5B model that runs on a 24GB graphics card. HunyuanVideo-1.5 is the pick if you have less memory, with a listed minimum of 14GB.
Can I use open source AI video commercially?
Often, but read each licence. Wan2.2 and Mochi 1 state Apache 2.0. LTX-2's code is Apache 2.0 while its models use the LTX community licences, and CogVideoX's 5B model has its own licence.
Can I run an open source video model on a Mac?
ComfyUI's repository lists support for Apple silicon and a desktop app for macOS. Whether a given model runs well depends on its memory needs, so check the model's own page.
How much GPU memory do I need?
From 4GB for CogVideoX's 2B model with optimisations, 14GB for HunyuanVideo-1.5 with offloading, and 24GB for Wan2.2's 5B model. The largest models ask for 60GB to 80GB.
Is Easy Reels open source?
No. It is a paid desktop app, $39 once, that runs locally. It makes animated and narrated reels and not photorealistic footage.