Wan2.2-T2V-A14B

Wan2.2-T2V-A14B, published on 28 July 2025 under Apache 2.0, is the model that carried the mixture-of-experts idea from language models into video diffusion. The design is unusual and worth understanding, because it is not the usual routing of tokens to experts. Denoising a video runs from pure noise to a finished frame, and Wan2.2 splits that trajectory in two: a high-noise expert works the early steps, where the job is overall layout and motion, and a low-noise expert takes over the later ones, where the job is detail. The handover point is not a fixed step count but the signal-to-noise ratio of the sample, with the switch set at half of the minimum. Each expert holds about 14 billion parameters, so the model totals 27 billion while only 14 billion are active at any step - capacity grows, the cost of running it does not. In practice the model generates five-second clips at 480P or 720P from a text prompt, and Alibaba ships a matching image-to-video sibling, I2V-A14B, with the same architecture. Running it as published is a data-centre job: the reference command in the model card asks for a GPU with at least 80 GB of memory. Alibaba's own benchmark, Wan-Bench 2.0, put it ahead of leading commercial models on most dimensions at release - a claim from the vendor's own scoreboard, which is how it should be read. The reason this profile matters more a year later is what came after it. Wan2.2 is the last numbered generation of the line whose weights were published at all - the served versions that followed, 2.5 through 3.0, are closed. The base itself kept moving in the open, though: the same account released Wan-Dancer in July 2026 and the Animate-2 character-animation weights in July and August 2026, all built on Wan2.2. What Alibaba closed is the flagship, not the whole output of the team. Everything Alibaba has served since - 2.5, 2.6, 2.7 and Wan3.0 - is API-only, and the newest files on the company's Hugging Face account are refinements of this model rather than successors to it. Anyone who wants a Wan they can run, inspect or fine-tune still ends up here.

▸Videos

No videos yet.

▸Reviews