I see that “Seedance 3.0” is being searched like it is a thing you can open right now, and every second AI video site has a page for it, I opened eight of those pages. Every one of them is a “Coming Soon” tile with a signup box under it. Not one has a model behind it, because ByteDance has not announced one. No model page on seed.bytedance.com, no API name, no date, nothing.
GOOGLE trends SEEDANCE 3.0 hit 76.
As someone who runs these clips for product cards and has to explain to clients why the third camera lens appeared, I wanted to know what is real, what is rumour, and what a 3.0 would even be given what 2.5 just did. So this is that, with dates, and then a test at the end on the version you can actually use today.
What Is Actually Out, With Dates
Instant answer on 3.0, NO. Here is the family as of today, 9 September 2026:
- Seedance 1.0, June 2025
- Seedance 2.0, 10 February 2026. Fifteen second clips, up to 9 image and 3 video/audio references, native audio, multi shot in one output
- Seedance 2.0 mini, 15 June 2026
- Seedance 2.5, previewed at the Volcano Engine FORCE conference in Beijing on 23 June 2026, rolling out from 31 July
- Seedance 3.0, nothing. Not on a roadmap, not on a waitlist, not teased
Two big steps between February and July. That cadence is the only hard fact anyone has about 3.0, and it is why every reseller has a page ready. If the gap holds, late 2026 or early 2027. If it does not hold, nobody knows, including the people selling you the waitlist.
Now I am not here to say resellers are lying, that is not the agenda. Seedance 3.0 on seedvideo.io is honest about it, the page says 2.0 is what runs today and 3.0 is coming, which is the correct way to write it. The problem is the sites that copy a spec sheet from a Reddit thread and format it like a datasheet.
What 2.5 Changed, Because That Is The Only Map We Have
If you want to guess a 3.0, you read what the team chose to fix between 2.0 and 2.5. They did not chase resolution. They chased three things:
- Length. 15 seconds went to 30 seconds in a single take, no stitching, and there is a 180 second beta floating around. Multi round extension on top of that for longer stories.
- References. 2.0 took about 12. 2.5 takes 50, split as 30 images, 10 video clips, 10 audio clips. ByteDance’s own demo had more than ten actor references in one generation and each one stayed where it was for the full 30 seconds.
- Editing instead of regenerating. Change one region of a finished clip and the rest stays put. Timestamp level edits. And the odd one, you can hand it an untextured 3D blockout, a clay model basically, plus material references, and it renders the moving shot to that shape.
Smaller stuff: native audio in 10 plus languages, their own claim of 20 percent better prompt adherence, 4K on paper.
So the direction is obvious. Longer, more things held consistent, and fixing a clip rather than rolling the dice again. Nothing in 2.5 says “prettier”.
![image: 2.0 vs 2.5 side by side, clip length, reference count, editing]
So What Would Actually Be In 3.0
This is my read, not ByteDance’s. Four things, in order of how sure I am.
1. Characters and products that survive between jobs. Right now the 50 references hold a face inside one 30 second clip. Close the job, start a new one tomorrow, upload the same 30 images again. Kling already has “Elements” for saving a subject, Runway Gen-4.5 has reference driven characters. ByteDance is behind on this one specific thing and they know it. A saved product or actor you call by name across separate generations is the most likely 3.0 headline. For anyone doing product cards this is the whole game, one approved chassis, reused for every colourway.
2. Longer, but not 18 minutes. The Reddit rumour says 10 to 18 minutes of continuous generation. Every model in this class, Seedance, Veo, Kling, Sora, all of them, is a diffusion transformer, and the attention cost in those goes up with the square of the length. That is why everybody’s number sits at 30 seconds and the 180 second thing is a beta. 60 or 90 seconds native in 3.0 I would believe. 18 minutes in one pass would need a different architecture, not a bigger one, and if they had that it would not leak on Reddit first.
3. Dubbing, yes, because it is half done already. The other rumour is native multilingual dubbing. 2.5 already does audio in 10 plus languages and 2.0 already led on lip sync, phoneme level. Taking an existing clip and re-voicing the mouth in another language is the natural next step from region editing plus audio references. I would put money on this one.
4. A draft mode that is close to real time. Lightricks’ LTX-Video 2 runs on a consumer GPU in near real time at lower quality. Directors want to see the move before they spend the credits. A fast, rough preview tier, then a full render, fits the “edit not regenerate” direction and the fact that 2.0 mini exists at all.
And one thing that will get tighter, not looser. After the February clips of real actors and the letter from two US senators in March asking ByteDance to shut Seedance down, the filters on real people and franchise characters moved to the model level. 3.0 ships with more of that. If your plan for 3.0 is celebrity ads, your plan is wrong on any version.
The Rumours, Sorted
- 10 to 18 minute continuous generation: possible in direction, unlikely in size, see above
- Native multilingual dubbing: likely, most of the parts are in 2.5
- 8K, “photoreal physics on par with Sora”: nobody serious is claiming this, it is filler on spec-sheet pages
- Release date: none. Any date you see is a guess dressed up
What I Tested On The Version You Can Use Today
Reading is fine, but I run this stuff for phone listing cards, so I did the cheap test on seedvideo.io with 2.0. Their homepage example is a 720p, 16:9, 5 second clip for 30 credits. That is not a hero shot, it is the “does the chassis hold” check, and it is all you need to learn most things.
I used an approved still of a two camera phone with a printed 128GB badge on the box in the shot, and asked only for a slow orbit. Audio off.
- The orbit itself is good. Better than Kling 2.6 on the same still, less wobble on the edges of the aluminium.
- Second three, as the phone turns, a third circle shows up on the rear. This is the thing every model does and it is why you count lenses before and after. Not a 2.0 problem, an everybody problem.
- The 128GB badge on the box went soft by second four and I could read it as 126 if I wanted to. Numbers in a clip are not safe on any model. Keep the number in the HTML.
- Cropping the box out of the still before generating fixed the badge issue entirely. Obvious, but nobody does it.
![image: frame at second three showing the extra lens, next to the original two lens still]
Am I saying do not use it? No. I am saying use it for motion on a frozen still and never for a fact. That advice does not change with 3.0, because the reference system is what got better in 2.5, and references are exactly how you stop the third lens. More references, better hold. 3.0 will most likely make that hold survive across jobs. It will not make a generated number trustworthy, no model has.
What To Do Now Instead Of Waiting
- Build your reference packs today. 30 images per product, every angle, the box, the ports. That pack is what you will feed 3.0 on day one and it is what makes 2.5 usable now.
- Run the 30 credit 5 second check on every still before you spend on a hero render. Count lenses at start, middle, last frame.
- Keep RAM, storage and price as typed text on the page. The clip moves the phone, it does not describe it.
- Do not pay anyone for “3.0 early access”. There is nothing to access. Paid exports on seedvideo.io are watermark free on 2.0 today, and when 3.0 lands it will show up on the same credit system, that is how every reseller works.
Go as per your ease on the platform. Just do not go as per a spec sheet that nobody at ByteDance has published.
