Wan 3.0 vs Wan 2.7: four models became one, and a cheap tier appeared
The headline upgrade is thirty seconds. The one that changes your monthly bill is 480P, a tier Wan 2.7 never had. Every number here is from Alibaba's own table.

If you built anything on Wan 2.7, the version number undersells what changed. The headline is duration — fifteen seconds to thirty. The structural change is that four separate models collapsed into one. And the change most likely to show up on your invoice is neither: Wan 3.0 introduced a 480P tier, which Wan 2.7 never had at any point, at a quarter of the 1080P rate.
Every figure below comes from Alibaba's own model tables, read on 2026-08-25. The comparison pieces currently ranking for this question are, in several cases, describing features that do not appear in either parameter table, so there is a corrections section near the end.
Both generations are documented in Alibaba's Model Studio: the Wan 3.0 API reference, and the video generation guide that still carries the 2.7 suite. Read 2026-08-25.
Four endpoints became one
Wan 2.7 shipped as a suite. You picked the model before you picked the shot:
| Wan 2.7 | Does | Duration | Resolutions |
|---|---|---|---|
wan2.7-t2v | Text to video | 2–15 s | 720P, 1080P |
wan2.7-i2v | First frame, first-and-last frame, continuation | 2–15 s | 720P, 1080P |
wan2.7-r2v | Reference images and video, voice reference | 2–15 s, 2–10 s with a reference video | 720P, 1080P |
wan2.7-videoedit | Instruction-based editing, migration | 2–10 s | 720P, 1080P |
Wan 3.0 replaces all four with wan3.0-video, plus a second tier,
wan3.0-video-prime. One endpoint, 2–30 seconds, 480P / 720P / 1080P, 30
fps, MP4.
That sounds like an implementation detail and it changes how you plan. On Wan 2.7 the mode was a commitment made before the prompt: choosing reference-to-video meant living inside ten seconds, and switching to keyframes meant a different model with different limits. On Wan 3.0 the input decides the mode and the ceilings stay the same either way.
One inheritance survives that consolidation, and it is the rule people trip over. Wan 3.0 accepts reference material or first and last frames, never both in the same request — the two families are mutually exclusive, and mixing them fails before generation starts. The modes merged into one endpoint; they did not merge into one job. The failure taxonomy is here.
Duration: fifteen to thirty, with one new constraint
Thirty seconds in a single continuous pass is the marquee number, and it is real: Wan 3.0 generates the whole clip in one generation rather than stitching shorter ones. At 30 fps that is 900 frames without a seam.
Two things about it that the announcement coverage tends to skip.
The floor moved too. Both generations start at two seconds, but on Wan 3.0 the 2–30 range applies to every mode, where Wan 2.7's reference and editing paths stopped at ten.
And there is a new ceiling that only appears when you feed video in. Input video duration plus output duration must total thirty seconds or less. Attach twelve seconds of reference footage and your maximum output is eighteen, not thirty. Requests past that are rejected outright, and the rule applies even when you hand duration control to the model.
The upgrade nobody is writing about: 480P
Wan 2.7 offered 720P and 1080P. That is the complete list, across all four models, in Alibaba's table.
Wan 3.0 adds 480P at $0.05 per second — a quarter of the 1080P rate. For anyone iterating rather than delivering, that single row is worth more than the extra fifteen seconds, because the expensive part of AI video was never the final render. It was the four attempts before it.
The arithmetic, ten seconds of output, four attempts:
| Best available route | Cost | |
|---|---|---|
| On Wan 2.7 | 4 × 720P, the cheapest tier there was | $4.00 |
| On Wan 3.0 | 3 × 480P drafts, then 1 × 1080P final | $3.50 |
You end up with a 1080P deliverable for less than four 720P attempts used to cost, and you saw three drafts along the way. Nothing about that requires the thirty-second headline. It requires a tier that did not exist eight months ago.
There is no quality catch worth worrying about here, either: 480P is the same Wan 3.0 model at fewer pixels, so framing, motion, pacing and sound all read truthfully. The full cost picture is here.
Genuinely new: documents and web pages as input
Wan 3.0 accepts a document or a URL as the source material, which no Wan 2.x model did.
| Limit | |
|---|---|
| Formats | docx, doc, xlsx, xls, pptx, ppt, pdf, txt, key, pages, numbers, md |
| Per request | 1 file or 1 link, never both, never several |
| Size | 100 MB |
| Pages | 50 |
Worth setting expectations: Wan 3.0 reads the document and directs a video from what it learned. It does not turn slide three into shot three. Fifty pages into thirty seconds is 0.6 seconds a page, so what comes back is closer to a trailer than a page-turn, which is the right instinct for a launch clip and the wrong one for a training module.
The reference budget roughly quadrupled
Wan 2.7's reference path took a handful of images or videos. Wan 3.0 takes 10 reference images, 5 reference videos and 5 reference audio clips in one request — twenty assets — with video and audio each capped at 15 seconds in total, and each asset addressable in the prompt as "Image 1", "Video 2" and so on.
One billing note that carries straight over into how you compose a request: Wan 3.0 bills reference video seconds at your output rate, while reference images, audio, documents and links add nothing. Ten stills are free; a fifteen-second reference clip is the most expensive thing you can attach.
What did not change, and what got more complicated
Still closed. Wan 3.0 has no published weights, no repository and no ComfyUI node, exactly like 2.7, 2.6 and 2.5 before it. Alibaba's last open-weight video flagship remains Wan 2.2. If self-hosting is the requirement, this upgrade is not available to you at any price — 2.2 is still the newest downloadable Wan, and the nearest current-generation model with published weights is MiniMax H3.
Frame rate did not move. Both generations output 30 fps. This is worth
saying because several comparison tables give Wan 3.0 as 24 fps, which would be
a downgrade if it were true. It is not; the value is in the API response under
usage.fps on both.
wan2.7-videoedit is still there. The editing model was not retired when
Wan 3.0 launched — it remains listed, at 720P and 1080P, up to ten seconds. Wan
3.0 does editing too, so for most work the newer model is the answer, but a 2.7
pipeline does not break on migration day.
There is a second price tier now. wan3.0-video-prime sits above the
standard model and costs meaningfully more per second. Alibaba publishes the
model ID and the price step but not a capability comparison, so anyone telling
you exactly what the premium buys is guessing. We are not going to guess either.
Six claims about this upgrade that the parameter tables do not support
Stated neutrally, because several of them appear on pages that rank well for this exact question. Each is checkable in one click.
| Circulating claim | Alibaba's documentation |
|---|---|
| A resolution tier above 1080P | There is no 4K tier. The Wan 3.0 resolution list is 480P, 720P, 1080P |
| Either version has open weights | Neither does. Open weights stop at Wan 2.2 |
| Wan 3.0 runs at 24 fps | 30 fps, in usage.fps, on both generations |
| A six-shot director mode with per-shot parameters | No such field exists in the parameter table |
| Cross-session character locking | No such field exists |
| Twelve reference assets, as nine images plus three videos plus three audio | 10 images, 5 videos, 5 audio — twenty, with different per-type caps |
The pattern in that list is worth noticing on your own behalf, not just ours: five of the six describe features so specific that they would need a named API parameter, and none of them has one. When a spec table gives a capability no request can express, the table was not written from the documentation.
Should you move
Yes, more or less immediately, if you iterate a lot. 480P changes your unit economics on day one, and nothing else in the migration requires effort — same async submit-and-poll pattern, same task states.
Yes, if your work needs one continuous take past fifteen seconds, or your source material is a deck, a report or a page rather than a prompt. Both are things Wan 2.7 cannot do at all.
Not urgently, if your output is short social clips at 720P and your existing prompts are tuned. The two-to-fifteen band works the same way it did, the frame rate is unchanged, and a tuned prompt library has real value. Move when you next have a reason to touch that code.
Not at all, if you need to run the model yourself. That door did not open, and there is no indication it is going to.
Questions
Is Wan 3.0 just Wan 2.7 with longer clips?
No. The duration ceiling doubled, but the bigger changes are the four-model suite collapsing into one endpoint, documents and web links as input, a 480P tier that did not exist, and a reference budget of twenty assets.
Do my Wan 2.7 prompts still work?
Broadly yes — the prompt grammar carries over and Wan 3.0's ceiling is far higher at 20,000 characters. The rewrite worth doing is structural: a prompt tuned for a ten-second shot rarely has enough beats to fill thirty seconds without drifting.
Does Wan 3.0 support 480P?
Yes, and Wan 2.7 never did. It is the cheapest tier at $0.05 per second, a quarter of 1080P, and it is the same model — which makes it the right place to do your first three attempts.
Was Wan 2.7 open source?
No, and neither is Wan 3.0. Four flagship generations in a row have shipped closed. Alibaba's last open-weight video model is Wan 2.2, still downloadable under Apache 2.0.
Can I still use wan2.7-videoedit?
It remains listed at 720P and 1080P up to ten seconds. Wan 3.0 handles editing as well, so new work should start there, but an existing 2.7 editing pipeline does not stop working the day you migrate everything else.
The one thing no table can tell you is whether a 480P tier changes how you work. Run the same prompt at both ends and compare the draft with the final — on Wan 3.0 the draft costs a quarter of it, and the first clip is free.
Written by
Editorial desk
wan-3.run


