Best Open-Source AI Video Models in 2026

The best open-source AI video model for most people in 2026 is still Wan 2.2: it is Apache 2.0, makes 480p and 720p clips, and you can run it on VdoBloom without a GPU at 36 credits for 5 seconds. HunyuanVideo 1.5 is the lightest strong option to run at home, from about 14 GB of video memory. LTX-2 adds synced audio, but its weights come under a community license rather than a classic open-source one.

Quick facts

• Open-weight model you can run on VdoBloom: Wan 2.2 A14B (text or image to video, 1 to 5 seconds, 720p, 36 credits for 5 s).
• Wan 2.2: Apache 2.0. HunyuanVideo 1.5: 8.3B parameters. LTX-2: LTX-2 Community License.
• Newer Wan versions (2.5 and later) and SkyReels V4 are hosted models, not open weights, as far as we could find in November 2026.
• On VdoBloom you pay per clip, with no subscription and credits that never expire.

Run Wan 2.2 in your browser

What "open source" means here

For video models, the useful question is: can you download the weights and run them yourself? We call those open-weight models. The license then decides what you may do with them. Apache 2.0 allows commercial use with few strings. A community license may add limits, such as revenue caps or use rules. Always read the license file for your use case. A model can be open to download and still restrict what you do with the output, so the licence matters as much as the weights.

How we ranked them

1. License. How freely you can use the weights and the output.
2. Quality. Motion, detail and prompt following.
3. Hardware. The smallest GPU that can run it.
4. Features. Image to video, audio, length and resolution.
5. Hosted access. Can you use it without your own GPU?

The models, ranked

1. Wan 2.2 (on VdoBloom): our pick and where we'd start

Alibaba's Wan 2.2, released under Apache 2.0, remains the most capable open-weight family. Its GitHub page lists text-to-video and image-to-video A14B models, which use a mixture-of-experts design with about 27B parameters in total and 14B active per step, at 480p and 720p. A smaller TI2V-5B model makes 720p video at 24 fps and is aimed at consumer GPUs. There are also speech-to-video and Animate models in the same repository.

On VdoBloom you can run Wan 2.2 A14B from the text to video or image to video tab, 1 to 5 seconds at 720p, 36 credits for 5 seconds. No download, no GPU, no setup. Try this prompt.

Animate a photo with Wan 2.2 for 36 credits

2. HunyuanVideo 1.5: best to run on a home GPU

Tencent's HunyuanVideo 1.5 is an 8.3B-parameter model that generates at 480p or 720p, with an optional super-resolution step to 1080p. Its GitHub page gives a minimum of 14 GB of video memory with model offloading turned on, which puts it within reach of high-end consumer cards. Read its license file before commercial use. It is not hosted on VdoBloom.

3. LTX-2: best open-access model with audio

Lightricks' LTX-2 generates video with synchronised audio and has pipelines for audio-driven video and lip-matched dubbing. The code is Apache 2.0, but the model weights are under the LTX-2 Community License and gated on Hugging Face, so you must accept the terms first. Treat it as open access rather than fully open source. It is not hosted on VdoBloom.

4. SkyReels V1 to V3

Skywork published open weights for earlier SkyReels versions. SkyReels V4, which VdoBloom hosts as SkyReels V4 image to video, did not have public weights that we could find as of November 2026. Try SkyReels V4.

Comparison table

Other open models. Older open video models such as Wan 2.1 are still used, mainly because many community tools and fine-tunes were built on them. They are useful for experiments but are behind Wan 2.2 on quality. New open releases arrive often, so check each project's GitHub page for the current version, licence and hardware notes before you plan a project around it. When a page and a third-party summary disagree, trust the project's own page.

Open weights vs hosted models

Self-hosting gives you privacy, fine-tuning and no per-clip cost once the hardware is paid for. It also means a large GPU, slow renders and your own updates. Hosted closed models such as Wan 3.0, Seedance 2.5 and Kling 4.0 are usually ahead on quality and length today; Wan 3.0 runs up to 30 seconds. A common setup is to prototype on open weights locally and render finals on a hosted model. On VdoBloom both kinds sit in the same model picker.

Here is a real Wan 3.0 clip, the model's own preview in the VdoBloom picker. Its original prompt was not recorded, so the prompt below describes the shot.

Prompt: "A hummingbird with shimmering green feathers hovers above a glowing orange flower, wings blurred, soft bokeh forest background, vertical close-up". A 5-second 720p run costs 32 credits. Recreate this.

Should you self-host?

Running an open video model yourself makes sense in a few cases. You need the footage to stay on your own machines. You want to fine-tune on your own characters or style. Or you will render so much that a GPU pays for itself. Outside those cases, hosted access is usually simpler.

• Hardware: the largest open models need data-centre GPUs. Even the lighter ones need a recent card with a lot of video memory, plus fast storage for the weights.
• Time: a short 720p clip can take many minutes on a consumer GPU. Hosted services run on bigger hardware.
• Setup: you manage Python environments, drivers, model downloads and updates.
• Licence checks: each model has its own terms, and some need you to accept them before download.
• Cost math: count the clips you expect to render in a month. At 36 credits per 5-second Wan 2.2 clip, a few dozen test clips cost far less than a new GPU, and you can stop at any time.
• Updates: new versions land often. A hosted picker gets them without a reinstall.

If you only want to see what Wan 2.2 can do, try it hosted first. A 5-second test on VdoBloom costs 36 credits and tells you whether the model suits your project before you set anything up.

How to run Wan 2.2 on VdoBloom

1. Open image to video (or text to video).
2. Pick Wan AI 2.2 A14B in the model list, or use the prefilled link above.
3. Upload a start image or type your prompt.
4. Choose a length of 1 to 5 seconds. The credit cost shows before you generate.
5. Generate and download the 720p clip.

Try hosted Wan 3.0 for 30-second clips

FAQ

What is the best open-source AI video model?

Wan 2.2. It is Apache 2.0, strong on motion and prompt following, and has both large and consumer-sized models.

Is Wan 3.0 open source?

We found no public Wan 3.0 weights as of November 2026. VdoBloom offers it as a hosted model.

Can I run an open video model without a GPU?

Yes, through a hosted service. On VdoBloom, Wan 2.2 A14B runs in the browser for 36 credits per 5-second clip.

How much GPU memory do I need?

HunyuanVideo 1.5 lists 14 GB with offloading. The Wan 2.2 5B model targets consumer GPUs; the A14B models need much more.

Can I use open-source video output commercially?

It depends on the license. Apache 2.0 is broadly permissive; community licenses may add conditions. Read the license file.

Related reading: Wan AI review 2026, Where to use Wan online, Best AI video models in 2026.