Which AI video models take an audio file as a reference on Sume?

On Sume, the video rows that accept audio as a reference are the Seedance 2.x models, Wan 3.0, MiniMax H3 and MiniMax H3 Max. gemini-omni-flash-1.1 takes images and video but no audio, and kling-3 takes no references at all.

MiniMax describes H3 as taking text, images, video and audio as one context (H3 post, read 2026-10-07); Sume's side is from its video docs and Video Router docs, read the same day.

Reference types by row

The H3 audio rules

Sume documents these H3 limits for audio references: up to 3 audio files, each 2 to 15 seconds, with a combined length of at most 15 seconds, and an audio file cannot be the only reference. Images, videos and audios together are capped at 12.

• Add at least one image or video alongside any audio.
• Trim each audio clip to 15 seconds or less before you upload.
• Keep the combined length of audio references within 15 seconds.
• Remember that H3 produces its own native stereo track, so a reference is guidance, not a fixed soundtrack.

What an audio reference is good for

MiniMax lists audio among the inputs H3 takes as context, so a reference is guidance for a new clip. If you need a recorded track to play as-is under the picture, a reference is the wrong tool; add the track in a timeline step after the clip is generated.

Choosing between the rows

Take minimax-h3 for a lower-cost 768p draft ($0.075 per second on Sume), minimax-h3-max for 1080p, and Wan 3.0 when you need more than 15 seconds. All three are priced as provider list times 1.25.

When no row takes audio

If you need to drive a clip with audio and your chosen row does not take it, generate the picture first and add the sound in assembly. On gemini-omni-flash-1.1 and kling-3 that is the route, since neither accepts an audio reference. Read the row's page in the docs each time, because the list of accepted inputs is per row and can change.

Checking the row before you build

The docs list the accepted inputs for each row, and the list is the only reliable source. Read it, send one small request with the audio attached, and look at the job result before you design a pipeline around the feature. A draft at the cheapest setting is enough to confirm that the audio changed the outcome. If it did not, you have spent cents rather than a project.