MiniMax H3 on Sume: generate_audio false is rejected, audio stays on

If you send generate_audio: false to minimax-h3 or minimax-h3-max on Sume, the request fails with model minimax-h3 always produces native stereo audio; omit generate_audio. Both rows generate sound and picture together and have no switch. To get a silent file, generate normally and mute the audio track afterwards, in your own editor or on a Sume timeline.

This differs from rows that do have an audio option, so a shared client should check the capability first.

What MiniMax says about audio

MiniMax's H3 announcement describes native stereo sound generated jointly with the video. That fits the Sume behavior: the sound is part of the generation, not a layer that can be left off. The practical outcome is that you pay for the clip with sound included, whatever you plan to do with it.

Ways to a silent result

None of these change what the model generates. They change what you deliver.

• Mute the track in your editor when you import the MP4.
• Build the final cut on a Sume timeline and leave the source audio out of the mix.
• Use a row where the audio flag is optional if the sound really must not be generated.

Guard your batch

Pipelines that set generate_audio on every request will fail only on these rows. Strip the field for them instead of relying on the 400. This helper drops it by model id.

NO_AUDIO_SWITCH = {"minimax-h3", "minimax-h3-max"}

def clean(body: dict) -> dict:
    if body.get("model") in NO_AUDIO_SWITCH:
        body = {k: v for k, v in body.items() if k != "generate_audio"}
    return body

print(clean({"model": "minimax-h3", "prompt": "x", "generate_audio": False}))

Sources and next reads

The row limits live on the Video Router page, and the job lifecycle is in Jobs and results. The vendor wording is in MiniMax's H3 announcement.