Seedance 2.5 vs MiniMax H3: both launched July 31, which to pin

Seedance 2.5 and MiniMax H3 (Hailuo 3.0) both launched on 2026-07-31, so people searching "Seedance 2.5 vs MiniMax H3" are comparing two models that arrived together. The short answer: Seedance 2.5 is the long-clip, many-reference model (up to 30 seconds per generation, per ByteDance Seed), while H3 is the shorter-clip model (4 to 15 seconds per a third-party summary). On Sume you can pin either by id and choose per job.

The Seedance numbers below come from the ByteDance Seed launch page (read 2026-10-05). The H3 numbers come from a third-party Hugging Face write-up (read 2026-10-05), not from MiniMax, so treat them as secondary. The Sume rows come from the Sume docs.

Vendor limits side by side

Reference counts are the clearest difference. Seed says Seedance 2.5 takes up to 30 images, 10 video clips and 10 audio clips in one pass. The third-party H3 summary lists 9 reference images, 3 video clips and 3 audio clips.

The same two models on Sume

Sume lists both in the Video Router catalog. Per the Video Router docs, seedance-2.5 accepts 4 to 30 seconds at 480p, 720p or 1080p, and minimax-h3 accepts 5 to 15 seconds at native 480p or 768p (768p is first-class, not 720p). There is also minimax-h3-max, the faster 768p variant, which does 480p, 768p and 1080p for 5 to 15 seconds, where 1080p is a latent refinement of native 768p.

The Sume docs say limits differ by model and tell you to read capabilities from GET /v1/video-router/models. Do that before you hard-code a reference count, because the vendor number and the number a hosted route accepts are not guaranteed to match.

Pick by the job

• One continuous shot longer than 15 seconds: seedance-2.5. H3 stops at 15 seconds.
• Many reference photos for one product or character: start with seedance-2.5, then check its capabilities for the count Sume accepts.
• Short hooks at 768p: minimax-h3 or minimax-h3-max; the Sume docs describe 768p as native.
• Same prompt on both for a taste test: send two requests with different model values and compare the outputs. Read each model's seed capability field in the catalog before you expect repeatable takes.

A pinned request

This is the documented Video Router request shape with the model pinned. Change model to minimax-h3 and duration to a value from 5 to 15 for the H3 run.

curl -X POST https://api.sume.com/v1/video-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: seed-vs-h3-001" \
  -d '{
    "model": "seedance-2.5",
    "prompt": "A vertical UGC-style product clip on a desk, natural light",
    "resolution": "720p",
    "duration": 12,
    "aspect_ratio": "9:16",
    "mode": "async"
  }'

Mistakes to avoid when you compare them

Same-day launches invite quick verdicts, and most of them are wrong in one of three ways. First, comparing a hosted 768p H3 clip to a 1080p Seedance clip and calling the gap a model difference when it is a resolution difference. Match resolution, or compare at the lowest resolution both support. Second, judging on a single take. These models are sampled, so run each prompt at least twice. Third, mixing reference sets: if one run gets nine images and the other gets three, you are comparing inputs, not models.

• Hold prompt, aspect ratio, duration and resolution constant.
• Log the job id of every take so stills can be pulled later.
• Write the decision rule before you look at outputs, for example subject consistency first, motion second.

Sources: ByteDance Seed, Hugging Face summary of H3, Sume Video Router docs.