How long can one AI video request be? Sume duration windows

Duration windows differ by model: 2 to 30 seconds on wan-3.0, 4 to 30 on seedance-2.5, 3 to 10 on Gemini Omni Flash 1.1. Read them from the catalog.

5 min readSume
All posts

Duration limits are per model on Sume. wan-3.0 accepts 2 to 30 seconds, seedance-2.5 accepts 4 to 30, minimax-h3 and minimax-h3-max accept 5 to 15, seedance-2 accepts 4 to 15, and gemini-omni-flash-1.1 accepts 3 to 10. Each other catalog model tops out at 15 seconds. Read the live values from GET /v1/videos/models.

The windows

The table collects the windows from the Sume video docs. It is a snapshot, not a contract. The catalog endpoint returns supported_durations for each model, and that is the value to code against.

A request outside the window is refused at submit. It does not run and trim. The one deliberate exception to a hard window is an edit with video_url: the source clip sets the output length, and any duration you send is only a reserve hint, 8 seconds by default.

From the Sume video docs, read 2026-10-05
Model idSeconds per requestNotes
gemini-omni-flash-1.13 to 10Native synced audio, 16:9 or 9:16
wan-3.02 to 30Widest short end of the catalog
seedance-2.54 to 30480p, 720p, 1080p
seedance-24 to 15Also 1080p
minimax-h35 to 15Native 480p and 768p
minimax-h3-max5 to 15Stereo audio, 480p to 1080p
higgsfield-genjutsu4 to 30Motion transfer, only when its provider is configured
h3-max-recast5 to 30One source video, 1 to 4 person photos

Read it from the catalog

Call the catalog before you submit, and validate your duration in code. This gives your app a clear message instead of a failed job.

curl "https://api.sume.com/v1/videos/models" \
  -H "Authorization: Bearer $SUME_API_KEY"

When your ad is longer than one request

Split the ad at scene boundaries, generate each clip, and join them. Timeline 1.0 assembles up to 200 video slots against one audio spine, and bills $0.10 per ceil output minute. Tell the model what the shot before and after looks like, or pin a last frame with frame_images so that two clips meet on the same picture.

Longer single clips are not always better. A 30-second request is one long wait and one failure point. Three 10-second jobs run in parallel up to your plan's concurrency.

Sources

Related posts

More in Models

All Models posts

Written by Sume