Seedance 2.5 is the stronger first test for a long, reference-heavy scene that would be expensive to split and reassemble. MiniMax H3 is the stronger first test for a 4–15 second audio-video shot, a hosted route to 2K, or a team that can genuinely benefit from controlling Base weights.
That recommendation is conditional, not a visual-quality verdict. A provider showcase, a third-party generation page, a BytePlus API result, MiniMax’s hosted pipeline, and a locally served checkpoint are different production surfaces. If those surfaces are not named, a comparison can look precise while answering the wrong question.
For a US team, there is an even earlier gate: the current H3 Community License excludes the United States. Do not plan a local H3 deployment until separate authorization is in place.
The decision in one production memo
| If the non-negotiable is… | Test first | Why |
|---|---|---|
| A continuous scene longer than 15 seconds | Seedance 2.5 | The current BytePlus route accepts 4–30 second output |
| Many character, scene, motion and audio references | Seedance 2.5 | The provider documents up to 30 images, 10 videos and 10 audio clips |
| Repairing a defined time range without rebuilding the entire concept | Seedance 2.5 | ByteDance describes timestamp-level editing and multi-round extension |
| A short clip with native stereo sound | MiniMax H3 | H3 documents 4–15 seconds and 32 kHz stereo output |
| A provider-supported 2K result | MiniMax H3 hosted workflow | H3-Regenerate-2K regenerates the Base result with the original context |
| Local control of the generative Base | MiniMax H3, only after legal and infrastructure review | H3-Base checkpoints are downloadable; the complete system is not fully local |
The table is a filter, not a winner board. It excludes candidates that cannot satisfy a hard delivery requirement. It does not prove which model will preserve your actor, product lettering, motion physics, or shot rhythm.
Two systems, not two single model files
ByteDance’s Seedance 2.5 announcement describes a managed audio-video model built around longer storytelling, mixed references, extension and editing. A single generation can reach 30 seconds. One request can include as many as 30 images, 10 video clips and 10 audio clips.
The practical benefit is fewer seams. A 24-second product story can stay in one generation instead of becoming two clips with a visible identity shift, a new sound bed and a repair pass at the join. More reference capacity also lets a team separate constraints: a character sheet for identity, a location video for camera language, product images for geometry and an audio clip for timing.
Capacity does not guarantee obedience. Conflicting references can make a larger request worse, and a 30-second output can accumulate more errors than a tightly scoped 8-second shot. The official claims establish what can be requested; your footage establishes what can be delivered.
MiniMax’s H3 model card describes a three-part system:
- H3-Context-IR interprets and restructures complex multimodal instructions.
- H3-Base generates video and audio, with a default 768-pixel short side.
- H3-Regenerate-2K feeds the Base result and original context back through H3 to produce the 2K result.
The Base checkpoints are available for local serving. Context-IR and Regenerate-2K are not included in the open release; MiniMax exposes API-backed workflows for them. “Open weights” therefore means control over the Base, not a self-contained copy of the complete official 2K experience.
That distinction matters in procurement. A local 768 route and the hosted 2K route have different data paths, credentials, latency, bills and operational dependencies. They should be separate rows in a test plan.
What the current output contracts actually say
The current BytePlus LAS enhanced-video documentation lists the model ID dreamina-seedance-2-5-260628. It supports 4–30 seconds, synchronized audio, common aspect ratios, and 480p or 720p output. The same page says 1080p is not supported for Seedance 2.5 on this route and that 4K is supported only by Seedance 2.0.
That is a route-specific fact. A consumer application or reseller may upscale, encode differently, or expose editing that LAS does not. A page advertising “4K Seedance 2.5” does not change the native BytePlus contract unless it explains the extra processing.
H3 documents 4–15 seconds, 24 FPS and 32 kHz stereo audio. Its Ref2VA mode accepts up to 9 images, 3 video clips and 3 audio clips, with no more than 12 mixed files. The Base result uses a 768 short side by default; the separate regeneration stage produces 2K.
Neither number alone predicts a better final export. Seedance can provide more time and references in one managed request. H3 can provide a higher-resolution hosted path for a shorter shot. The right unit of comparison is a usable deliverable, not maximum duration or pixel count in isolation.
US licensing changes the H3 branch
The MiniMax H3 Community License, dated August 2, 2026, defines the European Union, United Kingdom, Republic of Korea and United States as excluded territories. The license says use, reproduction, modification, distribution, running or display of H3 Works outside the applicable territory is not authorized under that agreement; MiniMax provides a path to request a separate license.
It also requires prior written authorization when commercial products and services generate more than USD 20 million in annual revenue. Distribution and hosted services carry additional notice, user-terms and safeguard obligations.
This is not a minor footnote after model evaluation. For a US company, it can eliminate the local-weights lane before a GPU is provisioned. A globally distributed product also needs review beyond the server location because the agreement discusses display and outputs. Have counsel read the current text and any separate authorization; this page is not legal advice.
Using a MiniMax consumer app or official hosted API is a separate service decision. Check that product’s regional availability and terms instead of assuming the community license is the only contract involved.
Cost: compare completed shots, not promotional clip prices
BytePlus publishes a calculable Seedance 2.5 formula on the LAS route. The enhanced-video base rate is USD 0.303 per second, multiplied by billable duration and a resolution factor. For a request without input video, Seedance 2.5 uses a factor of 0.6785 at 480p and 1.525 at 720p. With input video, billable time includes input plus output and the factors are 0.406 and 0.9125.
A successful 10-second, 720p request without input video is therefore estimated as:
0.303 × 10 × 1.525 = USD 4.62075
This is not a universal Seedance price. It is one documented BytePlus route, and the values can change. MiniMax’s current public pricing overview does not provide a first-party H3 per-generation figure that can be aligned to the same duration, resolution, audio and reference contract. Claims that H3 is a fixed percentage cheaper are not source-safe without a current quote or bill.
The useful cost denominator is not “one request.” Use cost per accepted shot:
accepted-shot cost = generation charges + failed attempts + regeneration or upscaling + audio repair + editing labor + self-hosting operations
If a USD 2 request needs four attempts and forty minutes of cleanup, it may lose to a USD 5 request accepted on the first pass. Record actual charges and labor for the same deliverable.
Run a three-shot acceptance test
A fair test does not require identical prompt text. Each model has different timing and reference syntax, so forcing the same string can favor whichever parser happens to understand it. Keep the creative contract constant and adapt the instructions to each official interface.
Use three shots that expose different failure modes:
- A sequence longer than 15 seconds with two motivated camera changes. Measure identity, scene continuity, audio continuity and whether the ending resolves the planned action.
- A short product shot containing fine text, a hand interaction and a material change. Measure usable product geometry and lettering, not only atmosphere.
- A close dialogue shot in the delivery language. Measure pronunciation, lip timing, voice consistency, accidental captions and repair time.
For every shot, record attempts to first acceptance, usable seconds, instruction completion, identity or product retention, audio usability, wall-clock wait, manual repair and the invoice. Keep provider post-processing visible in the notes. A hosted H3 2K result must not be labeled as a local Base result, and a third-party Seedance upscale must not be labeled as native LAS output.
The result usually produces a clean choice:
- Choose Seedance 2.5 when the shot cannot be reduced below 15 seconds without expensive seams, when many references must work together, or when targeted editing removes more re-generation than it adds.
- Choose hosted H3 when short native-audio shots and the 2K regeneration route beat longer continuity in value.
- Consider local H3-Base only when authorization, infrastructure, moderation and maintenance are already viable business capabilities.
- Choose neither when both models repeatedly damage the product, actor or required action; change the shot design or expand the model shortlist.
If you still need a market-wide shortlist, use the 2026 AI video model map. If Seedance is already selected and the remaining decision is 2.5 versus 2.0, Fast or Mini, the Seedance model family guide is the narrower next step.
As of August 26, 2026, the practical split is stable: Seedance 2.5 is the managed long-form and deep-reference lane; MiniMax H3 is the short native-audio, hosted-2K and controllable-Base lane. The model that wins your project is the one that produces the accepted shot with the lower total production cost—after the route and license are made explicit.
