How many reference video seconds? Kling 4.0, Wan 3.0, H3, Omni

How much reference video can each model take? Kling 4.0 allows up to 5 videos totaling 30 seconds on its page, MiniMax H3 up to 3 videos and 15 seconds, Wan 3.0 on Sume up to 5 videos and 15 seconds, and Omni up to 3 clips of 3 seconds. Of these, Wan, H3 and Omni are callable on Sume; Kling 4.0 is not in the Sume catalog.

Reading the numbers

The Kling 4.0 post lists up to 15 total assets, 10 images, 5 videos, up to 7 subjects and voice references, and says its full launch is October 2026 with no API details. The MiniMax guide adds up to 9 images, 3 audio clips and 12 files combined. The Omni docs cap video references at 3 clips of up to 3 seconds.

Why it matters

A 15 second source clip fits H3 and Wan but not Omni, which would need to be cut to 3 seconds. A 30 second performance reference only fits Kling 4.0 as written. If you need to reuse a long clip on Sume, trim it first with video-trim and send the section that carries the motion.

Sume's catalog notes also differ in kind: Omni takes images and videos but no audio, while Wan and MiniMax take audio references too.

Before you send

Confirm the live fields in GET /v1/videos/models; a model accepts a reference type only when supported_input_references lists it, and an unsupported one is rejected. Count seconds on the trimmed file, not the original.

Trimming reference clips

On Sume, POST /v1/video-trim cuts a range from a workspace clip into a new MP4 for $0.02 per job, and the result can be sent as a reference. It requires start, one of end or duration, and an Idempotency-Key. Use precision: exact for a frame-accurate cut, and the cut will be a fresh re-encode rather than a stream copy. For Wan, remember the frame-rate floor of 16 fps on reference videos, so check the source before you rely on it.

Numbers for Seedance 2.5 and Kling 3 on Sume are not in the table because I did not read per-type video reference limits for them; Kling 3 has no reference URL fields in the catalog at all.