The short version
- Seedance 2.5 wins the spec sheet — 4K vs 2K, 30 seconds vs 15, 50 reference assets vs 12
- Both are live on Creaa.ai now, on the same unlimited plan — so you can pick per shot instead of per subscription
- H3 shipped open weights and a hosted API in the same week, at roughly a third of mainstream pricing
- H3 ranks first for video editing on Artificial Analysis, edges Seedance 2.0 on text-to-video, and trails it on image-to-video
- H3 lists at $0.13/s at 2K, $0.09/s at 768p — but the same model costs up to 2× more depending on which reseller you buy it through
- Rule of thumb: reach for Seedance 2.5 when you need length, resolution or many references; reach for H3 when you are editing footage or starting from text
Both are live — so pick per shot, not per subscription
Every comparison written in the first week put Seedance 2.5 on top of the spec sheet. Native 4K, 30-second single-shot generation, up to 50 reference assets — it's a genuinely impressive jump over 2.0. That part is still true.
What changed since those comparisons were written is availability. Seedance 2.5 launched inside ByteDance's own Jimeng app first, with no third-party access; MiniMax went the other way, publishing H3's weights on Hugging Face alongside a hosted API. For a few days the practical answer really was "use H3, because it's the one you can call."
That gap has closed. Both models are live on Creaa.ai now, and both sit on the same unlimited plan. You don't have to bet a subscription on one of them — you can switch models between shots and let each do what it's actually good at.
Which makes the rest of this comparison more useful than it would have been a week ago. When you can only get one model, the choice is made for you. When you can get both, the question becomes specific: what does each one win at?
Spec-by-spec
| MiniMax H3 | Seedance 2.5 | |
|---|---|---|
| Availability | Open weights + hosted API, both live | Live on Creaa.ai; ByteDance keeps the weights closed |
| Max resolution | 2K (1440p) | Native 4K |
| Max duration | 15 seconds | 30 seconds, single shot |
| Reference assets | 12 | 50 |
| Native audio | Yes — 32 kHz stereo | Yes |
| Frame rate | 24 fps | 24 fps |
| Architecture | 33B dense single-stream Omni-Transformer | Not disclosed |
| Run it yourself | Yes — 4×H200 recommended; 2×RTX 5090 works with offload | Not possible |
| Licence | Community License — commercial use under $20M revenue, attribution required | Closed, commercial terms |
Specifications as published in early August 2026. Both vendors iterate quickly — check current documentation before committing to either.
What the benchmarks actually say
Spec sheets and output quality aren't the same thing. On Artificial Analysis, the picture is more interesting than "Seedance wins":
- Video editing — H3 ranks first. That's the strongest single result either model has, and it's the use case least discussed in the coverage.
- Text-to-video — H3 (1242) edges out Seedance 2.0 (1225). Note the comparison is against 2.0, not 2.5; independent 2.5 numbers aren't available yet, for the same reason there's no API.
- Image-to-video — H3 trails both Seedance 2.0 and Gemini Omni Flash.
Open weights vs closed: what it changes
The open-weight release isn't just a philosophical difference. It has three concrete consequences that show up within weeks:
- Price competition. Once weights are public, multiple inference providers host the same model and compete. That drives per-second cost down fast — the same pattern every open-weight release has followed.
- No single point of failure. A closed model can change price, change terms, or restrict access, and you have no alternative. With open weights you can move providers, or run it yourself.
- Nothing gets deprecated out from under you. This year has already shown what that's worth — a major closed video model was announced, then retired, with no successor path for anyone who had built on it.
Against that, closed models tend to ship the frontier first. Seedance 2.5's 4K and 30-second generation are real advantages that no open model currently matches.
Should you self-host H3?
You'll see "it runs on two RTX 5090s" repeated a lot this week. That's true, but it's the footnote, not the headline. SGLang's own cookbook lists a 2×RTX 5090 recipe that only works with layerwise offload of the DiT, text encoder and VAE — and the throughput figures published alongside it are from a 5-step run, an order of magnitude below the standard 50-step config. The recommended default is 4×H200. Two consumer cards will let you evaluate the model; they will not let you serve customers with it.
There's also a licence boundary worth reading before you build on it. The MiniMax H3 Community License permits commercial use for organisations under US$20M annual revenue, with prominent attribution required. Above that threshold you need a separate agreement with MiniMax. Most teams are comfortably inside it — but the attribution requirement applies to everyone, and it's the kind of thing that's easier to design in than to retrofit.
| Self-host makes sense when… | Hosted API makes sense when… |
|---|---|
| You generate consistently, most hours of the day | Your load is bursty or unpredictable |
| Data can't leave your infrastructure | You're still evaluating quality |
| You have someone who can own GPU operations | You'd rather not run inference infrastructure |
| You need a fixed cost ceiling regardless of volume | You want provider price competition to work in your favour |
The cost picture
MiniMax's own list price is $0.13 per second at 2K and $0.09 per second at 768p (768p is still gated at the time of writing). That's roughly a third of what mainstream closed video models charge. Seedance 2.5 is priced above 2.0, which is the normal pattern for a new frontier release.
What's less obvious is that the same model costs twice as much depending on where you buy it. H3 is one model with one set of weights, but the resale spread is wide:
| Where you buy it | 768p | 2K |
|---|---|---|
| MiniMax, direct | $0.09/s | $0.13/s |
| WaveSpeedAI | $0.09/s | $0.13/s |
| Vercel AI Gateway · OpenRouter | $0.09/s | $0.13/s |
| Atlas Cloud | $0.10/s | $0.14/s |
| fal | — | $0.26/s |
List prices observed early August 2026. MiniMax also publishes a separate RMB rate card on its domestic platform, which converts to a lower figure than the international one — worth checking if you can bill in CNY.
Three billing details that don't show up in the headline rate, and that catch people out:
- Reference video is billed by its own input duration, at the output resolution's rate. A 5-second clip with a 3-second reference video bills as 8 seconds. If reference-to-video is your main workflow, your real rate is well above the sticker price.
- Reference audio is free, and so are the first five reference images. Extra images carry a small per-image fee.
- Some resellers impose a minimum billable duration — a 4-second floor turns every short generation into a 4-second charge.
Even so, the number that decides your bill isn't price per second — it's cost per finished clip. Most creators run three to ten takes before a shot is usable. A model that's 30% cheaper but needs twice as many retakes is more expensive, and no pricing page will tell you that. Only your own prompts will.
How to choose
| Your situation | What to do |
|---|---|
| You're building on an API | H3. Seedance 2.5 isn't an option yet, and Seedance 2.0 remains the closed alternative |
| Your work is image-to-video | Stay on Seedance 2.0 for now — it's ahead of H3 on exactly that task |
| You do a lot of video editing | H3. It's the current leader in that category |
| You need 4K or 30-second shots | Neither is available to you outside ByteDance's app. Wait for the 2.5 API |
| You just want volume without counting credits | Cost per finished clip matters more than the model name — see our pricing breakdown |
Where we stand: Creaa.ai runs Seedance 2 today. We'll add Seedance 2.5 when ByteDance ships a public API — we can't offer what doesn't exist yet, and we'd rather say that than imply otherwise. If you generate at volume and want to stop counting credits per clip, that part works now.
FAQ
MiniMax H3 or Seedance — which should I use?
It depends on the shot, and on which Seedance you mean. Against Seedance 2.5, H3 loses on length (15s vs 30s), resolution and reference count, but costs roughly a third as much. Against Seedance 2.0, H3 scores slightly higher on text-to-video and ranks first in the world for video editing — but loses on image-to-video, which is what most people actually do. Short version: reach for Seedance when you're animating a still or need length and resolution; reach for H3 when you're editing or restyling footage, or starting from text.
Can I use Seedance 2.5 outside ByteDance's own app?
Yes. It launched inside ByteDance's Jimeng app first with no third-party access, but that gap has closed — Seedance 2.5 runs on Creaa.ai now, on the same unlimited plan as MiniMax H3. ByteDance still keeps the weights closed, so unlike H3 you can't self-host it.
Are the MiniMax H3 weights actually released?
Yes — published on Hugging Face under the MiniMax H3 Community License. It's a 33B dense single-stream Omni-Transformer. A hosted API is live on fal if you'd rather not download anything.
Which is better overall?
Seedance 2.5 on paper; H3 in practice this month, because you can actually use it. If you work image-to-video, note that H3's weakest benchmark result is exactly that task.
Can I run H3 locally?
Yes, but read the fine print. SGLang documents a 2×RTX 5090 recipe that requires layerwise offload; the recommended default is 4×H200. Two consumer cards are enough to evaluate the model, not to serve traffic with it. Validate quality through the hosted API before buying anything — and check the Community License's $20M revenue threshold and attribution requirement first.
Does H3 generate audio?
Yes, 32 kHz stereo generated natively with the video. Worth checking on any model you migrate to: without native audio, every clip quietly gains a post-production step.
Keep reading
Specifications, pricing and availability observed in early August 2026 and subject to change — verify with each vendor before committing. Benchmark references are to Artificial Analysis as reported at launch. Creaa.ai has no affiliation with MiniMax or ByteDance; we provide access to Seedance models and describe competing models as accurately as we can, including where they outperform what we offer.