⚔️ MiniMax H3 vs Seedance 2.5: Open Weights vs Native 4K

ByteDance and MiniMax both launched on July 31, and Bloomberg wrote them up as one story. They are not the same kind of release. MiniMax shipped a documented public API on day one. ByteDance shipped into two consumer apps and said the API is coming.

That asymmetry runs through everything below, including how much of this comparison can even be sourced: MiniMax publishes a full API reference, while ByteDance has published no resolution spec, no pricing and no model IDs for 2.5.

By the numbers

	MiniMax H3	Seedance 2.5
Launched	July 31, 2026	July 31, 2026 (announced June 23)
Max resolution	2K native	Not published by ByteDance
Frame rate	24 fps fixed	Not published
Single-shot length	4–15 seconds	Up to 30 seconds, extendable twice
Native audio	Yes, same pass, no toggle	Not a headline feature
Weights	Open, announced not shipped	Closed
Reference inputs	9 images + 3 video + 3 audio, 12 max	30 images + 10 video + 10 audio, 50 max
Editing	Not on the v2 API	Timestamp-level, green-screen, camera re-edit
3D pre-visualization	No	Yes (white-model blockout)
API status	Live	"Coming soon" via BytePlus ModelArk
Published price	~$0.13/sec (third-party)	None published

The 15-second wall is the real decision

Everything else is negotiable. This is not.

H3 caps a single generation at fifteen seconds, and the API rejects anything above it. Seedance 2.5 runs to thirty in one continuous pass and can be extended twice beyond that. If your shot is a character crossing a room, delivering a line, and reaching a window — one move, no cuts — H3 physically cannot do it and no prompt trick changes that.

The counter-argument is that H3 can cut inside its fifteen seconds. It will render an intro, a highlight and a close as three shots in one generation. That buys you screen time without buying continuity, which is a different product. A three-cut fifteen-second sequence is not a thirty-second take.

Kling 3.0 Omni sits in the same bracket from a third direction — up to six cuts inside 15 seconds at native 4K — which we broke down in the Omni guide and measured against ByteDance in Kling 3.0 vs Seedance 2.5.

Audio is where H3 takes it back

Seedance 2.5's headline features are resolution and length. H3's is that the file comes back with sound in it — spoken dialogue timed to the mouth, effects hit on the frame, room tone underneath.

For dialogue coverage that difference is an hour of post per shot. You write the line in the prompt, in quotes, and skip the voice model and the waveform-nudging entirely — the full H3 guide covers the prompt structure that makes the audio land on the right frames. For silent establishing shots it is worth nothing at all, and you are back to comparing resolution.

Veo 3.1 and Sora 2 also generate native audio, so H3 is not alone here; what it adds is voice carry-over from a reference clip and open weights.

Open weights, and what that actually buys

MiniMax shipped H3 as an open-weight model. ByteDance kept Seedance 2.5 closed.

The practical difference shows up in three places. Client work under NDA where footage cannot cross a third-party server. Long-run cost, where a rented GPU beats per-second API billing past a certain volume. And permanence — an API tier can be repriced or retired, a downloaded checkpoint cannot.

Set against that: the weights are announced but not shipped as of August 1 — MiniMax says "within days" under its Community License — nobody has credible consumer-GPU VRAM figures, and a 2K video model is not going to be kind to a 12GB card. For what actually runs on consumer hardware today, see open-weight AI video models. Open weights are a real advantage on a six-month horizon and close to zero this week.

Quality: no clean winner

Artificial Analysis puts H3 first in Video Editing and top-three in both text-to-video and image-to-video. On its text-to-video-with-audio board, Gemini Omni Flash leads at Elo 1246 with H3 second at 1242; on image-to-video-with-audio, Seedance 2.0 leads at 1196, Gemini Omni Flash follows at 1195, and H3 sits third at 1185.

Seedance 2.5 is not on those boards yet — the 2.0 entry is what ranks. ByteDance claims roughly a 20% prompt-adherence gain over 2.0, and no independent benchmark has tested it. Both models are days old and these placements will move.

Decision table

Your shot needs	Take
Spoken dialogue with matching mouth	H3
One continuous take past 15 seconds	Seedance 2.5
4K master, no upscale	Neither is confirmed — ByteDance has published no 2.5 resolution spec
Footage that never leaves your hardware	H3
Camera and lighting planned before render	Seedance 2.5 (3D previz)
Voice or character locked across many shots	H3 (reference mode)
On-screen text, logos, packaging	Neither — both still fail this

How we picked

Both models are days old and this is a specification comparison, not a shootout — we have not run matched footage through either yet. Every H3 figure is read off MiniMax's published v2 API reference rather than the launch blog, which is where the hard limits came from. Seedance 2.5 specs are read off ByteDance Seed's own launch post, cross-checked against CineD. We have corrected an earlier version of this page that repeated a pre-launch "native 4K" figure from June reporting — ByteDance's actual announcement publishes no resolution spec at all. Details in what ByteDance has confirmed. The open-versus-closed framing is from SCMP and TechTimes reporting on the July 31 releases.

We have deliberately not published a cost-per-finished-shot comparison. That requires matched outputs on both models, and the pricing to compare them against does not exist in public yet.

What didn't make the cut

Raw ELO leaderboard positions. Both models are inside their first week, community rankings are thin, and the numbers will not survive the month.

We also skipped the "which is better for TikTok" angle. At 9:16 and eight seconds, both clear the bar and the choice comes down to whether the clip talks.

The honest summary

If you shoot dialogue, H3 pays for itself in post time and the 2K ceiling rarely bites at social resolutions. If you shoot long unbroken takes, Seedance 2.5 is the only one of the two that reaches thirty seconds — though you are buying into a model whose resolution, price and API date ByteDance has not published. Most production slates contain both, which is the split we are planning around — H3 for coverage, Seedance for the shots that have to breathe.