Kling 4.0 Omni Reference: 15 inputs; kling-3 on Sume takes none

Kling 4.0 Omni Reference combines images, videos and subjects in one generation, up to 15 combined assets. On Sume, kling-3 takes no references at all (its catalog constraint is no reference_*_urls). Ids that do take multi-type references include wan-3.0, minimax-h3 and gemini-omni-flash-1.1; the table also lists the retiring Video 1.0 alias.

Kling facts are from its blog post, where the full model is "coming in October 2026"; Sume facts are from the API catalog and Video 1.0, read 2026-09-30.

What does Kling 4.0 Omni Reference accept?

The post lists up to 10 images, up to 5 videos with a combined duration of up to 30 seconds, and up to 7 subjects (including up to 3 video-based subjects), within a combined total of 15 reference assets. It adds that individual input limits apply inside that total.

Which Sume ids take references, and how many?

The caps below come from catalog constraints. They are per-type caps; the catalog states a combined cap only for minimax-h3.

Are there combined limits I should know?

For minimax-h3 the catalog says images, videos and audios together may not exceed 12, and audio cannot be the only reference. For the other ids the catalog states per-type caps only, so a Kling-style single total of 15 does not map across.

What should I do if my workflow assumes kling-3 references?

Pick one of the ids above and send reference_image_urls, reference_video_urls or reference_audio_urls on POST /v1/video-router/generate. Read capabilities from GET /v1/video-router/models first, as the Video Router docs advise. For Wan details see Wan 3.0 reference limits.