Best open source AI video models in 2026 (and which ones aren't)
Wan 2.2, LTX-2.5, MiniMax H3, HunyuanVideo 1.5, LongCat-Video: licenses read on October 25, 2026, excluded countries, revenue caps and stated hardware.
Frank HoubreFounder of Imaginode
8 min read
Open source AI video models: the short answer
As of October 2026, five video model families publish their weights: Wan 2.2, LongCat-Video, LTX-2.5, MiniMax H3 and HunyuanVideo 1.5. Only the first two use a standard open source license (Apache 2.0 and MIT). The other three are downloadable under "community" licenses that cap your revenue or leave out whole regions. MiniMax H3 excludes the United States and the EU; HunyuanVideo 1.5 excludes the EU, the UK and South Korea.
So if you want a video model you can run on your own GPU and sell work from without a clause catching up with you later, the safe pick is Wan 2.2. LTX-2.5 is the most capable of the open models (sound generated with the picture, several shots in a row), and it stays free as long as your company makes under $10 million a year.
What about Wan 3.0, the one everybody is talking about? It isn't open source. No official repository publishes its weights; you rent it by the second through an API. Plenty of sites say otherwise, and we get to that below.
Disclosure: you're reading this on Imaginode, Frank Houbre's tool. Imaginode doesn't run video models on your machine. It offers MiniMax H3 as a hosted model paid in credits, and the cost section says so plainly.
Open source vs open weights: not the same deal
A truly open source model releases its weights under a license that lets anyone use, modify and sell with it, with no conditions on country or company size.
Apache 2.0 and MIT work that way. Keep the license notice and you're mostly done.
An open weights model also releases its files, but under a custom license, usually called a community license: free up to some threshold, with banned uses and sometimes a list of countries where it doesn't apply. The download button on Hugging Face looks identical. The difference sits in the LICENSE file, which almost nobody opens.
The last case is the sneakiest: a model that carries the name of an open family while its new version publishes nothing. Wan is the textbook example. Versions 2.1 and 2.2 are Apache 2.0; versions 2.5 through 3.0 are closed.
License table as of October 25, 2026
Every row comes from the model's license file or official card, read on the same day. "Not published" means the publisher gives no figure, and we don't make one up.
Open weights video models: license, commercial use and excluded regions
| Model | License | Commercial use | Excluded regions | Stated hardware |
|---|---|---|---|---|
| Wan 2.2 (Alibaba) | Apache 2.0 | Unrestricted | None | 24 GB VRAM (TI2V-5B, 720p) |
| LongCat-Video (Meituan) | MIT | Unrestricted | None | Not published (13.6B parameters) |
| LTX-2.5 (Lightricks) | LTX-2.x Community | Free under $10M annual revenue | Sanctioned countries | Not published |
| MiniMax H3 | MiniMax H3 Community | Written approval above $20M revenue | EU, UK, South Korea, United States | Not published (33B parameters) |
| HunyuanVideo 1.5 (Tencent) | Tencent Hunyuan Community | License required above 100M monthly users | EU, UK, South Korea | 14 GB with offloading |
| Wan 3.0 (Alibaba) | Closed, API only | Per the API terms | Not applicable | Nothing to install |
Wan 2.2: the no-strings choice
Wan 2.2 is the only major recent video model under Apache 2.0, and Alibaba states in writing that it claims no rights over what you generate.
The family has several models released since July 28, 2025: 14B text-to-video and image-to-video, a compact TI2V-5B that does both, a sound-to-video model (S2V), and Wan2.2-Animate-2, released in August 2026 to transfer motion from a video onto a character.
The compact one is what matters for a solo creator. According to its model card, it renders 720p at 24 fps on a 24 GB card such as an RTX 4090, and produces a 5-second clip in under 9 minutes without special optimization. Slow, sure, but it runs on your desk and costs you electricity.
Its weak spot is no secret: Wan 2.2 dates from summer 2025. Its text-to-video model has no native audio, and the closed models of 2026 clearly beat it on faces and long motion. When the license matters more than the latest look, it's still where you start.
LTX-2.5: the most capable, if you're small enough
Lightricks' LTX-2.5 generates picture and sound in one pass, chains several shots while keeping characters and setting, and reaches 4K through a bundled upscaler.
It's the open model closest to what the closed tools do. It takes text, an image or a video as input, and its model card shows examples at 1920 by 1088. The LTX-2 family also got one thing right that others skip: it works in ComfyUI on release day, with ready-made workflows.
The license is the LTX-2 Community License, dated January 5, 2026. Commercial use is free while your company, subsidiaries included, makes under $10 million a year. Above that, you need a paid agreement with Lightricks. No country is singled out apart from sanctioned ones, and Lightricks claims no rights over your videos.
Two things to know before you start. Downloading LTX-2.5 means agreeing to share your Hugging Face details, including consent to Lightricks' marketing. And the model card gives no VRAM figure, only options for smaller cards (fp8 quantization, CPU offload).
MiniMax H3: the most downloaded, and off-limits in the US
MiniMax H3 is the most downloaded open video model right now, over 5 million downloads last month on Hugging Face, and its license doesn't apply in the United States, the EU, the UK or South Korea.
On paper, H3 has everything: 33 billion parameters, 4 to 15 second clips at 24 fps, stereo sound, dialogue in 11 languages, image, video and voice references. MiniMax released it on August 2, 2026 under its own MiniMax H3 Community License.
Open the file. Section I.5 defines the "Excluded Territories": the European Union, the United Kingdom, the Republic of Korea and the United States of America. The license is granted only in the rest of the world. If your studio is in New York or Paris, the community license doesn't cover you, and you need separate authorization from MiniMax.
The rest is worth reading too. Above $20 million in revenue, you need written approval. Any commercial product using H3 must display "MiniMax H3" in its interface. And you can't use its outputs to improve another AI model. Also, the 2K version (H3-Regenerate-2K) isn't released, so the open weights stop at 768 pixels on the short side.
HunyuanVideo 1.5: light, but not in Europe
Tencent's HunyuanVideo 1.5 is one of the lightest recent models, 8.3 billion parameters and a 14 GB VRAM minimum with offloading, and its license also leaves out the EU.
It renders at 480p or 720p, and a second network upscales the result to 1080p. Along with Wan 2.2's TI2V-5B, it asks for the least video memory on this list, going by the publishers' own figures.
Its Tencent Hunyuan Community License defines the "Territory" as worldwide except the European Union, the United Kingdom and South Korea. Unlike H3, the United States is covered. Above 100 million monthly active users, you must ask Tencent for a license. And as with MiniMax, outputs can't be used to improve another AI model.
LongCat-Video: MIT, built for long clips
Meituan's LongCat-Video is MIT licensed, the most permissive on this list, and targets videos several minutes long.
The model has 13.6 billion parameters and handles text-to-video, image-to-video and continuing an existing video. Its card claims 720p at 30 fps and "minutes-long" videos without color drift. That's the publisher's claim, not a test.
Adoption is far lower: about 2,500 downloads last month, against more than 5 million for H3. Fewer ready-made workflows, less help on forums. But if the license is your first filter and you work with long shots, it's the only other truly free option next to Wan 2.2.
Is Wan 3.0 open source? No
Wan 3.0 and Wan 3.0 Prime are closed models: Alibaba sells them by the second through its Model Studio API and through hosts like fal.ai, without releasing weights.
The official Wan-AI account on Hugging Face, captured above, has no Wan 2.5, 2.6, 2.7 or 3.0 repository. Neither does the Wan-Video GitHub. Pages announcing "Wan 3.0 under Apache 2.0" mix up the family and the version.
On Alibaba Cloud's pricing page, read on October 25, 2026, Wan 3.0 costs $0.05 per second at 480p, $0.10 at 720p and $0.20 at 1080p (Singapore region, with a 30-second free quota and a limited-time 30% discount shown). For comparison, the API version of Wan 2.2 (wan2.2-t2v-plus) costs $0.02 per second at 480p.
Same story at MiniMax: H3 is open, but H3-Regenerate-2K isn't released, and H3 Max only exists as an API.
Run it locally or pay per second?
Local makes sense if you already own a 24 GB card and have lots of shots to render; otherwise renting by the second costs less than people assume.
Do the math with Wan 2.2's numbers. Under 9 minutes for 5 seconds of 720p on an RTX 4090 means roughly 20 minutes for a 10-second shot, before counting failed takes. For a 40-shot short with three takes per shot, the machine runs for days. If you already own the card, that's free. If you have to buy it, you need a lot of shots to pay it off.
Hosted, you pay per clip. In Imaginode, MiniMax H3 costs 78 credits for 5 seconds, about $0.94 according to its model page, and a failed generation is refunded in credits automatically. Nothing to install, and the service's terms apply rather than the weights license. The limits are real too: Imaginode caps H3 at 10 seconds per clip, and it doesn't run any video model on your own machine (its local AI connection only covers the Image node).
Our take: learn on a hosted model, keep Wan 2.2 locally for projects where you need to own the whole pipeline, and read the license before you deliver anything to a client.
Which open model for which job
You want zero legal risk for commercial work: Wan 2.2, or LongCat-Video if your shots run long.
You want sound and several consistent shots, and your company makes under $10 million a year: LTX-2.5, once you accept the access form.
You have a modest GPU and you're outside Europe: HunyuanVideo 1.5, which starts at 14 GB with offloading.
You're in the US or the EU and want MiniMax H3: the community license doesn't cover you. Use a hosted service or ask MiniMax for a license.
Before you judge quality, test your shortlist on your own shots with the same method for each: our guide on how to compare AI video models walks through it. And if you'd rather find the cheapest place to run a closed model, see where to use Seedance 2.5 or our pick of the best AI video generators.
How we analysed this
No video was generated for this article and no model was installed locally: it compares licenses and model cards, not outputs. The LICENSE files and model cards were read on Hugging Face and GitHub on October 25, 2026, logged out, and captured the same day.
Hardware figures are the ones the publishers give, when they give any. Wan API prices come from Alibaba Cloud's pricing page, read the same day. The cost of MiniMax H3 in Imaginode comes from its public model page. Licenses change: always reread the repository's LICENSE file before commercial use. This article is not legal advice.