Wan 3.0 is the 2026 flagship video model from Alibaba's Tongyi Lab, and unlike the Wan models that made the family famous, it is API-only: there are no published weights to download.
That single fact is the thing most write-ups get wrong, so it is worth stating up front. We have not generated with Wan 3.0 yet and it is not on the 8frame canvas, so everything below is split into what we verified ourselves and what Alibaba claims for the beta.
What is actually verifiable
No open weights. We queried the Hugging Face API for the Wan-AI organization on 2026-08-28. The newest published repositories are all in the Wan 2.2 line (the most recent upload being Wan2.2-Animate-2-14B-Distilled-Diffusers, 2026-08-06). There is no Wan 2.5 repo and no Wan 3.0 repo. You can run that check yourself in one request against huggingface.co/api/models?author=Wan-AI.
It is reaching Western infrastructure. Wan 3.0 first shipped in public beta on Alibaba's own cloud in early August 2026. In our weekly model-catalog sweep on 2026-08-28 we found alibaba/wan-3.0/* and a separate alibaba/wan-3.0-prime/* tier listed as active on fal, dated 2026-08-24. Per-second prices were not published on those pages at the time of the sweep, which is why we have not priced it into the canvas yet.
The open Wan line stopped at 2.2. Wan 2.1 and Wan 2.2 shipped with downloadable weights under Apache 2.0. Wan 2.5 was announced but never published weights, and Wan 3.0 continued that pattern. If someone tells you Wan 3.0 is open source, ask them for the repository link.
What Alibaba claims for the beta
The beta materials describe three headline changes over Wan 2.5: clips around 30 seconds generated in a single pass rather than stitched, output up to 1080p with synchronized audio produced in the same pass, and reference inputs that go beyond text and images to documents, spreadsheets, slides, PDFs, and web pages. That last one is the genuinely novel piece. Every other major model in 2026 takes text, images, and sometimes video as conditioning. Turning a deck into a video is a different input class.
We have not tested any of it, so treat those as vendor claims until someone runs a side-by-side.
Which Wan you can actually run today
| Model | Weights | Where it runs | 8frame price |
|---|---|---|---|
| Wan 2.2 | Apache 2.0, on Hugging Face | Your GPU, or hosted | 50 credits / 5s at 720p |
| Wan 2.5 | Not published | Hosted API only | 102 credits / 5s at 1080p |
| Wan 3.0 | Not published | Alibaba Cloud, appearing on fal | Not on the canvas yet |
Credits cost $0.01 each in a top-up pack, so a 5-second Wan 2.5 clip is about $1.02 and the 720p Wan 2.2 clip is about $0.50. The $19 Starter plan carries 1,000 credits a month.
Related concepts
What is Wan covers the family's history and why it anchored the budget end of the market. Is Wan 3.0 open source walks through the license question with the receipts, and best open source AI video generator 2026 covers what you can actually self-host now that the Wan line has gone closed above 2.2.
When Wan 3.0 gets published pricing we will run it against the same prompts as everything else and price it into the canvas. Until then, see how the models we do host compare on identical prompts in best AI video generator 2026.