MiniMax H3 audio references: 2 to 15 s each, 15 s combined

For minimax-h3 and minimax-h3-max on Sume, each audio reference must be between 2 and 15 seconds, all audio together can add up to at most 15 seconds, you can send at most 3 files, and audio cannot be your only reference. Send at least one image or video next to it.

A 1-second sound effect is too short, and a 40-second song needs trimming first. The numbers come from the Video Router catalog constraints on both H3 rows.

The full envelope

Audio shares a budget with the other reference types, so count everything.

Preparing an audio reference

Cut the clip to the part that carries the sound you want, between 2 and 15 seconds. If you start from a longer file, video_trim handles range cuts on a clip hosted on Sume, and the result is a new artifact you can pass as a URL. Keep the file on public HTTPS so the provider can download it.

If you have three voices or sound sources, keep each at 2 to 15 seconds and check that the three lengths add up to 15 or less. Two 8-second files already overshoot.

What the model still does on its own

H3 produces native stereo audio and has no generate_audio toggle, so a reference is a steer, not a switch that turns sound on. If you want a silent clip, pick an id that has the toggle instead. The H3 durations are 5 to 15 seconds, and Sume bills by output seconds at list price times 1.25, so adding references does not change the per-second rate, except for images past the fifth.

A practical rule: start with one image plus one 5-second audio sample, check the result, then add references one at a time. It is easier to see which reference changed the output.

Pre-flight list

Run these checks on each audio file before it becomes a reference.

Keep the originals of every trimmed clip. If a result is off, you can change the cut point and re-run without having to hunt for the source again.

• Length is at least 2 seconds and at most 15.
• The sum of all audio references is 15 seconds or less.
• No more than 3 audio files, and at least one image or video alongside.
• The whole reference set stays at 12 items or fewer, counting images and videos.