Wan2.2 split its video model into two paired experts

The shot

On 28 July 2025, Alibaba's Wan team logged in its own repository news feed: 'We've released the inference code and model weights of Wan2.2.' The GitHub repository describes Wan2.2 as 'a major upgrade' to the Wan2.1 open-weights model already covered on this site, built on a 'Mixture-of-Experts (MoE) architecture' that splits denoising across a high-noise and a low-noise expert model. The repository states training data grew by '+65.6% more images and +83.2% more videos' compared with Wan2.1, and that the models are licensed under 'the Apache 2.0 License.' The Hugging Face organization page lists the resulting repositories, including the T2V-A14B, I2V-A14B and TI2V-5B checkpoints named in the GitHub documentation.

What the documents show

The repository states Wan2.2 'achieves TOP performance among all open-sourced and closed-sourced models' on 'our new Wan-Bench 2.0,' a comparison the developer ran and names no specific competitor system for, unlike some other vendors in this cycle who name rivals directly. The TI2V-5B variant is documented separately as a '5B dense model' rather than a mixture-of-experts model, supporting 720p generation at 24 frames per second and able to produce, in the repository's words, 'a 5-second 720P video in under 9 minutes on a single consumer-grade GPU.' The Hugging Face listing corroborates that these named checkpoints exist as separate, independently downloadable repositories rather than a single bundled release.

The workflow

A production adopting Wan2.2 chooses among differently sized checkpoints by hardware budget: the A14B mixture-of-experts models support 480p and 720p generation for teams with larger GPU allocations, while TI2V-5B targets a single consumer card. The stated Apache 2.0 terms mean the license does not by itself restrict commercial use of generated output, though the repository is explicit that a user remains 'fully accountable' for how that output is used. None of this substitutes for a production's own review of a generated clip's continuity or its fit within a larger cut.

What the tool does not change

Alibaba's own benchmark claim on Wan-Bench 2.0 is the developer's self-reported result on an evaluation the same developer designed, not an outcome a third party replicated. The repository documents version-to-version change, Wan2.1 to Wan2.2, in its own words rather than describing every capability as new; a team should not assume every Wan2.1 limitation was solved simply because a newer version exists.

• Was a stated benchmark run on an evaluation the vendor designed, and does the source name any competitor system directly.
• Which checkpoint size and resolution combination actually matches the hardware a production has available.
• Does a licensing term address generated-output rights, or only the code and weights themselves.

Wan2.2's own repository documents a dated, versioned upgrade with a stated architecture change and self-reported benchmark, giving a production concrete grounds for comparison against Wan2.1 without treating the vendor's own performance claim as independently settled.