Open-weight video generation models

last edited

Open-weight video generation models include Wan2.2, released in July 2025. Another is MOVA, a joint video and audio model from early 2026. Open-weight releases in this area publish downloadable weights, code, and a permissive grant. This page covers architecture and tooling for open-weight video generation.

Wan2.2

Wan2.2 introduced a mixture-of-experts design in video diffusion. It upgrades the earlier 2.1 release. The public release came on July 28, 2025. The code is published in the Wan2.2 GitHub repository.

  • Commercial use: The weights are open and commercial use is allowed. The Wan2.2 model page lists the exact license.
  • Training data: 65.6% more images and 83.2% more video than the earlier release.
  • 5B model: The autoencoder uses 16×16×4 compression. It generates 720P video at 24fps and runs on a consumer RTX 4090.
  • ComfyUI: An official native workflow is documented in the ComfyUI Wan2.2 tutorial.

MOVA

MOVA is an end-to-end model for synchronized video and audio generation. The code reached GitHub on January 29. A technical report followed on February 10. Evaluation code was added on May 6. The project lives in the MOVA GitHub repository.

The model has 32B total parameters, with 18B active. The MOVA 720p model page lists an open license.

Related References

This page forms part of AI Video Generation — Organized Reference. Broader comparisons appear in Wan 2.2 and specialist models and in video model roundup. Specialized releases are listed in open-source and specialized models.

Sources