Official sources only

Wan 3.0 vs Seedance 2.5

Two 30-second models announced a week apart. Neither is open to everyone, and neither has published its weights.

This comparison sticks to four axes and quotes only what Alibaba and ByteDance have documented themselves — no aggregator figures on either side. Where a vendor has not published something, it is marked as such rather than filled in.

Neither model is available on ImagineToVideo

One of these models is now in our generator: Seedance 2.5 went live on ImagineToVideo on August 7, 2026 via its public API (FAL). Wan 3.0 remains invite-only with no public API, so you cannot generate with it here. This page compares them on the record — and where 2.5's claims have become verifiable, we say so.

Both sides last verified on

Where the two actually differ

Both target 30 seconds

Wan 3.0's API reference documents a 2–30 second range. ByteDance describes Seedance 2.5 as producing 30 seconds of audio and video in a single pass. On this axis the two claim the same ceiling.

Reference handling is the real gap

Seedance 2.5's announcement quotes far higher counts — up to 30 images, 10 video clips and 10 audio clips. Wan 3.0 documents seven input types instead, including files and links, but publishes no per-type limits.

Both are gated, differently

Wan 3.0 is an invite-only test on QwenCloud that you apply for. Seedance 2.5 now has a live public API (FAL) with published pricing — and it runs in production on ImagineToVideo, where you can generate with it today.

Neither has published weights

No Wan 3.0 repository exists on Hugging Face or GitHub as of the verification date, and ByteDance has not released Seedance weights either. Self-hosting is not an option for either model.

The four axes

Wan 3.0's column comes from QwenCloud's changelog and API reference. Seedance 2.5's column comes from ByteDance's own release announcement. Nothing here is sourced from a third party.

Wan 3.0Seedance 2.5
Maximum duration
30 seconds of audio and video in a single pass, per ByteDance's release announcement.

Vendor claim; not independently reproduced.

Reference material: types and counts
Up to 30 images, 10 video clips and 10 audio clips as reference, plus timestamp-level editing. ⚠️ Documents, spreadsheets and web pages are not among the stated types.

Counts come from the announcement; there is no public API reference to check them against.

How you get access
Are the weights available?
No. ByteDance has not published Seedance 2.5 weights, and Seedance has not been an open-weights line.

Deliberately four axes and no more. Adding rows would mean filling cells that neither vendor has published — which is how comparison pages start inventing numbers. Both columns are re-checked weekly.

Which one is worth tracking for your case

You work from a large reference library

Seedance 2.5 states the higher numbers by a wide margin — 30 images plus 10 video and 10 audio clips at the API level, with timestamp-level editing. Wan 3.0 accepts a broader set of input types, including files and links, but publishes no counts, so the two cannot be ranked on this axis yet.

You need something you can commit to

Only one is committable today: Seedance 2.5 has a live API with published token pricing and runs on ImagineToVideo. Wan 3.0 has a documented request schema you can design against while you wait for access, but no pricing and no way in without an invitation.

How this comparison was put together

So you can reproduce it, or catch us if a source moves.

1

Take each side's own sources

Wan 3.0 from QwenCloud's model release log and its wan3.0-video API reference. Seedance 2.5 from ByteDance's release announcement. No summary or aggregator page is used on either side.

2

Keep the axes to what both have published

Duration, reference handling, access and weight availability are the only four points where both vendors have said something on the record. Anything else would be one side's marketing against the other's silence.

3

Mark the gaps instead of filling them

Where a vendor has not published a figure — Wan's per-type reference limits, either side's pricing — it carries a ⚠️ rather than an estimate.

4

Re-check weekly

Both models are moving. Every source above is checked again each week, and any change is written into our Wan 3.0 page's update log the same day.

Wan 3.0 vs Seedance 2.5: common questions

Which is better, Wan 3.0 or Seedance 2.5?

Output-quality rankings would still be guesswork — no rigorous third-party benchmark exists. But the availability gap is now decisive: Seedance 2.5 has a live public API with published pricing and runs in production (including on ImagineToVideo), while Wan 3.0 remains invite-only. On documented capability, 2.5 states much higher reference counts; Wan 3.0 accepts a broader range of input types.

Can I use either of them today?

Seedance 2.5 — yes: it went live on ImagineToVideo on August 7, 2026 and is set as the default model in the generator. Wan 3.0 — no: it is an invite-only test on QwenCloud that you have to apply for. Our alternatives page lists everything you can run right now.

Which one generates longer video?

They claim the same ceiling. Wan 3.0's API reference documents 2 to 30 seconds; ByteDance describes Seedance 2.5 as producing 30 seconds of audio and video in one pass. The difference between them is not length.

Which supports more reference material?

By stated counts, Seedance 2.5 — up to 30 images, 10 video clips and 10 audio clips at the API level, with timestamp-level editing. On ImagineToVideo the launch integration exposes 9 reference images + 3 reference videos. Wan 3.0 documents seven input types including files and links, but ⚠️ publishes no per-type limits, so a direct count comparison is not possible yet.

Is either one open source?

No. As of August 7, 2026 there is no Wan 3.0 repository on the Wan-AI organisation on Hugging Face or the Wan-Video organisation on GitHub, and ByteDance has not published Seedance 2.5 weights. Self-hosting is not available for either.

What do they cost?

Seedance 2.5 now has published API pricing: FAL bills $0.0214 per 1,000 tokens — roughly $0.47 per second at 720p and $0.22 at 480p with audio. ⚠️ Wan 3.0 has none: QwenCloud's API reference for wan3.0-video lists no rate, and ByteDance's own Volcano Ark rate card also remains unpublished.

Both are gated. Some models are not.

If you need 30-second clips with sound and reference control now rather than eventually, there are models available on this site today — with their real parameter limits listed.

Wan 3.0 vs Seedance 2.5: A Sourced Comparison