Quick answer
Seedance 2.5 and MiniMax H3 launched the same day with opposite priorities. Seedance goes long: 30-second single-pass clips and 50 reference slots, at 480p to 1080p. H3 goes sharp and open: native 2K at 24fps, 15-second clips, 12 references, and open weights you can run yourself. Choose Seedance for duration and heavy reference control, H3 for resolution today or for self-hosting. Seedance has no native 4K tier despite widespread claims otherwise.
On 31 July 2026, ByteDance shipped Seedance 2.5 and MiniMax shipped H3. Same day, and almost opposite bets about what matters in AI video.
Seedance went long: 30 seconds in one pass, 50 reference slots, heavy control over what each input governs. H3 went sharp and open: native 2K, 15-second clips, and weights you can download.
Most comparisons of the two repeat a resolution claim that is backwards. That is the first thing to fix, because it is the one that will wreck a delivery plan.
The correction
Vendor pages widely describe Seedance 2.5 as scaling to native 4K. It does not, today.
- Released
- 31 July 2026. Jimeng and Doubao first, then Dreamina and third-party APIs.
- Clip length
- 4 to 30 seconds in one pass, or duration set to auto. Beta long-video toward ~3 minutes.
- Resolution
- 480p and 720p, with 1080p arriving around 17 August 2026. 4K is a post-generation upscale.Widely mis-stated: that it "scales to native 4K". An upscale is not native capture, and the difference shows on fine texture and text. (Disputed)
- References
- 50 total: 30 images, 10 video, 10 audio. Tagged @Image1, @Video1, @Audio1 by upload order.
- Audio
- Co-generated in the same pass. generate_audio defaults to true.
- Weights
- Closed. API and first-party apps only.
- Pricing
- Per second, and additionally scaling with reference count.
ByteDance launch materials plus provider documentation (fal, KIE, BytePlus, EvoLink). Model specs move quickly; re-check before committing a production budget.
Amber rows are ones where published sources disagree. We've said which reading we trust and why rather than picking one silently.
- Released
- 31 July 2026, after a WAIC preview on 17 July. Also called Hailuo 3.0 or Hailuo 03.
- Architecture
- 33B dense single-stream H3-Omni-Transformer, with multimodal rotary position embeddings.
- Clip length
- 4 to 15 seconds, extendable to roughly 30 seconds with the Extend tool.
- Resolution
- Native 2560x1440 at 24fps. 1080p often capped near 5 to 6 seconds; 768p to about 10 seconds.
- References
- 12 omni files across text, image, video and audio in any combination.
- Audio
- Native stereo, generated in the same pass.
- Weights
- Open. Self-hostable and fine-tunable, subject to licence.
- Pricing
- Roughly $0.0104/s at 512p, $0.04/s at 768p, $0.08/s at 1080p. Reference video adds billable input seconds; images and audio do not. (Unconfirmed)
MiniMax launch materials, the H3 model card, and provider documentation (KIE, EvoLink, APIMart). Reported rates vary by provider.
So on resolution the underdog currently wins, and it is not close: native 1440p against a model that is only now reaching 1080p. If you are delivering this month and pixel density is the constraint, that decides it.
Head to head
| Seedance 2.5 | MiniMax H3 | |
|---|---|---|
| Max single-pass length | 30s | 15s (~30s via Extend) |
| Native resolution | 480p / 720p, 1080p arriving | 2560x1440 @ 24fps |
| 4K | Upscale only | Not offered |
| References | 50 (30 img / 10 vid / 10 aud) | 12 omni, any mix |
| Native audio | Yes | Yes, stereo |
| Weights | Closed | Open |
| Prompt-strength control | None exposed | CFG scale (5.0 to 7.0 typical) |
| Billing | Per second, scales with reference count | Per second |
What each bet buys you
Seedance: duration and reference control
Thirty seconds in one pass is not just a longer clip, it is a different unit of work. A 30-second single take can carry a three-act beat structure without a cut, which is not something you assemble from two 15-second generations without visible seams. If your output is a narrative scene, that matters more than pixel count.
The 50-slot reference budget is the other half. Splitting references across 30 images, 10 videos and 10 audio clips lets you pin identity, wardrobe, environment, grade, motion and music separately, each with its own stated job. For work that puts a real person in a scene, that separation is the whole game, and 12 combined slots gets tight as soon as you have two characters and a product that also has to stay consistent.
The catch is that references are priced, so a heavy stack makes every draft more expensive. Cheap iteration means a lean stack.
H3: resolution, and the one that actually matters
Native 2K at 24fps is a real advantage today, with the caveat that the highest tiers come with shorter length caps: 1080p often tops out near 5 to 6 seconds, 768p around 10. So H3's resolution win is strongest on short-form work, which is also where most of it will be used.
But the consequential difference is open weights. Seedance is a closed API. H3 can be downloaded, inspected, self-hosted and fine-tuned, subject to its licence. If you have data-residency obligations, if you need the model to keep working regardless of a vendor's roadmap, or if you want to fine-tune on your own material, that single fact outweighs every row in the table above. No amount of prompt-adherence advantage substitutes for being able to run the thing yourself.
H3 also exposes a CFG scale, typically useful between 5.0 and 7.0, which trades literal prompt-following against creative interpretation. Seedance exposes no equivalent, so on Seedance that trade is made for you.
Prompting differences worth knowing
The grammar is broadly shared. Both want a subject, an action, camera language, lighting, and references with stated jobs. Two differences are worth carrying between them:
- Beat structure matters more on Seedance, simply because 30 seconds gives the model more room to wander. A 15-second H3 clip tolerates a continuous description in a way a 30-second Seedance clip does not. Writing 30 seconds covers the structure.
- Reference scoping matters more on H3, because 12 slots means each one is carrying more weight, and an unscoped reference bleeds more of itself into the output when there are fewer inputs to average against.
If you already write good Seedance prompts, your H3 prompts will be fine. The reverse is also true, with the beat caveat.
Which one to pick
Choose Seedance 2.5 if: you need more than 15 seconds in one take; you are putting real people in scenes and need to pin identity separately from wardrobe and environment; you have several characters or products to keep consistent; or you want the tighter prompt adherence for complex staging.
Choose MiniMax H3 if: resolution is your binding constraint this month; you are making short-form where 15 seconds is plenty; you need stereo audio; or you need to run the model yourself, for compliance, cost or independence.
Use both if you can. They are close enough in prompt grammar that carrying a shot between them is cheap, and having a second model is the fastest way to work out whether a strange result is a model problem or a provider problem. Kie.ai carries both through one API with per-second billing, which makes that comparison a one-line change rather than a second integration.
A caution about launch clips
Neither model should be judged from its launch reel. Those are curated from many attempts, at settings and prompt lengths nobody publishes, and both companies had the same incentive on the same day.
Run your own shot, on both, at the shortest usable length and lowest resolution. It costs a few cents and it will tell you more than any comparison table, including this one.
Where to go next
- Seedance 2.5 prompt handbook for the full grammar
- All 50 reference slots for the reference budget in depth
- Seedance vs Kling vs Veo for the wider field
- When Seedance ignores your prompt for the debugging side
We run Seedance in production at Starrd, which is why the Seedance column here is more opinionated than the H3 one. Where we have only read the documentation rather than shipped against it, the tables above say so.