The 30-second club just got its second member. ByteDance's Seedance 2.5 has been generating half-minute takes with synced audio for a while now. On August 6, Alibaba opened the public beta for Wan 3.0 and claimed the same headline number. Launch coverage points to a formal event on August 10 in Hangzhou, so by the time you read this there may be more official material than I had.
That timing gap shapes everything below. One of these models I can poke at through a third-party channel and describe from its actual parameter surface. The other is days old, lives behind Alibaba's own platforms, and has no public API documentation yet. I am not going to pretend those are symmetrical sources, so this comparison is explicit about where each fact comes from.
What each side can prove today
| Seedance 2.5 | Wan 3.0 | |
|---|---|---|
| Length | 4 to 30 s, you pick the exact second | "Native 30s Duration" with what Alibaba calls intelligent duration control |
| Audio | on by default, dialogue via double quotes, can be switched off | "sound design" and "immersive audiovisual" per the official site; no documented control surface yet |
| References | 30 images + 10 videos + 10 audio files | "up to 20 reference assets, including complex document and webpage parsing" |
| Resolution | 480p / 720p on available channels | not officially pinned; beta listings show a 480p tier and coverage reports 1080p |
| Editing | generation only; chaining via a returned last frame | instruction-based and reference-based editing, presented as a core feature |
| Access | third-party web platforms and API aggregators, now | Alibaba's own platforms, public beta; API "opening soon" |
| Open weights | no | no; the Wan-Video GitHub org still stops at Wan 2.2 |
Sources for the right column: the wan.video homepage copy, the Wan-Video GitHub organization, and the fact that Alibaba Cloud's Model Studio documentation index still tops out at Wan 2.7 API references as I write this. Everything on the left comes from using Seedance 2.5 and reading its request schema.
The reference systems solve different problems
This is the comparison that actually matters, and it is not about the numbers 50 and 20.
Seedance 2.5 references are media: you feed it images, clips, and audio tracks, and they answer the question "make it look and sound like this." A character sheet keeps a face stable across takes. An audio track sets the mood the video gets generated to fit. It is a consistency machine.
Wan 3.0's headline reference feature is document and webpage parsing. Feed it a deck, a spreadsheet, a page, and it builds a video of that content. That answers a different question entirely: "make a video about this." If the demos hold up, that is closer to an explainer-video pipeline than to a cinematic generator.
Those are both legitimate directions. They are just not the same product decision, and which one you need depends on whether your input is a mood board or a quarterly report. I have seen nothing yet on how well Wan 3.0 handles the media-consistency job, and Seedance has no answer at all to the document job.
Editing is Wan's bet
Seedance 2.5, at least through the channels I can reach, generates. If you want to change a finished clip you regenerate, or you chain a new segment off the returned last frame. There is no edit mode.
Wan 3.0 puts "Precision Video Editing" on its front page, with both instruction-driven and reference-driven variants. Editing inside the generation model, rather than in a separate tool, would collapse a real step out of the workflow. Whether it works well enough to rely on is exactly the kind of thing a homepage cannot tell you, and I flag it here as the feature I most want to test rather than one I can vouch for.
The openness twist
Wan built its reputation on open weights. Wan 2.1 and 2.2 are sitting on GitHub and a lot of self-hosted pipelines run them. So it is worth saying plainly: Wan 3.0 has no downloadable weights, no checkpoint, no repo. As of this week it is a hosted, closed model, same as Seedance 2.5. If you were waiting for 3.0 to upgrade a self-hosted stack, nothing has shipped for you yet, and nobody official has said whether it will.
What I cannot compare and will not fake
Output quality, because I have not generated with Wan 3.0 and a launch-week verdict based on cherry-picked demo reels would be worthless anyway. Pricing, because Seedance costs vary by channel and Wan's beta pricing is fragmentary. And the fine print that decides real integrations, like Wan's prompt limits, rate behavior, or how those 20 reference slots are actually typed, because the API docs that would answer this do not exist publicly yet.
That last gap cuts one way for now. If you need a 30-second audio-native model inside a pipeline this week, Seedance 2.5 is the one with a schema you can script against today; the parameter surface is all exposed in the browser at seedance 2.5 if you want to see it without writing code. Wan 3.0 deserves a real evaluation in about a month, once the documentation lands and the launch-week fog clears. I plan to write that follow-up, and I honestly do not know yet which way it will go.
Top comments (0)