Shortest AI video clip by model: MiniMax H3 4 s, Sume row 5 s

What is the shortest clip you can generate? It depends on the model, and vendor pages and Sume rows do not always match. MiniMax's guide lists H3 at 4 to 15 seconds; Sume's minimax-h3 row starts at 5. On Sume, Gemini Omni 1.1 Flash is the short-clip row at 3 seconds. If you need a shorter clip than a model's floor, generate longer and trim.

Minimum durations

Sume's numbers come from the catalog in the video models docs. Vendor numbers come from the pages cited in this post's sources.

Trimming on Sume

The POST /v1/video-trim endpoint takes a video_url, a start and either an end or a duration, and requires an Idempotency-Key header. The video-trim docs describe it. A typical flow is to render a 5 second H3 clip and trim it to 4, so you pay for the 5 second generation plus the trim, which the docs list at $0.02 per job. The clip must already be a media.sume.com artifact of your workspace, and the request needs an Idempotency-Key.

If the real goal is a very short, cheap clip, Wan 3.0's 2 second floor or Omni's 3 second floor is a better fit than trimming.

Before you build around a floor

Minimums are exactly the kind of number a vendor changes without announcement. Read supported_durations for each model from GET /v1/videos/models at request time rather than hard-coding, and let your code choose the smallest allowed duration at or above the length you want.

Why the gap exists

A catalog row is a promise Sume makes after its own testing and provider mapping, so it can be narrower than the vendor page. When the vendor allows 4 seconds and the Sume row says 5, request 5 and trim. When the vendor page is silent on the minimum, as for several rows above, the Sume row is the only number this post can confirm.

Also note that Kling 4.0 is in the table only as a vendor fact; it is not in Sume's catalog, and the Kling page gives no API details.