An independent browser interface for Wan 3.0, Alibaba's video model. Four inputs reach one endpoint: a prompt up to 20,000 characters; a photo as first and/or last frame; a reference set of 10 images, 5 clips and 5 audio tracks cited by number; or a document up to 50 pages the model reads and directs from. Out comes 2-30 whole seconds in one pass at 30 fps, speech and music alongside the picture. No cloud account, no API key, no GPU. Every result prints its model and task ID. First clip free.
I built this because reaching Wan 3.0 officially means an Alibaba Cloud account, a region choice, an API key that has to match that region, and an asynchronous polling loop — for a thing most people want to try for ten minutes.
The part that surprised me while building it was how much of the cost is decided by defaults rather than by prompting. If you do not set `resolution`, the API generates at 1080P: four times the 480P rate, no error, no warning. And there are two separate 24-hour clocks — one on the task ID, one on the output file — so a job can succeed and still leave you with nothing if no one collected it.
So the design goal ended up being honesty rather than magic: print the model ID and task ID under every clip, refund failures automatically, and keep a dated spec table with each figure linked to Alibaba's own documentation (including the four things write-ups get wrong: no 4K tier, 30 fps not 24, Wan 2.2 is the last open-weight flagship, and muting the audio does not make a clip cheaper).
Happy to answer anything about the model's actual limits.