MiniMax H3 limits: 9 images, 3 videos, 3 audio, file caps

MiniMax's video guide, read 2026-10-03, limits an H3 request to a 7,000 character prompt, up to 9 reference images, 3 reference videos and 3 reference audio files. File size caps are 30 MB per image, 50 MB per video and 15 MB per audio file. Sume's catalog entries minimax-h3 and minimax-h3-max carry their own duration and resolution limits, listed below.

The vendor limits

The guide separates H3 from H3-Max, which it lists with fal.ai. The numbers differ on duration and resolution, not on reference counts.

The Sume side

The Sume docs describe the catalog entries differently, and the two lists should not be merged. From the Video Router doc and the video generation guide:

• minimax-h3 accepts 5-15 s at native 480p or 768p. 768p is first-class, not 720p, and 2K or 4K upscales are priced when requested.
• minimax-h3-max is the faster 768p variant: text to video, first and last frame image to video, and reference to video at 480p, 768p or 1080p (1080p is a latent refinement from native 768p) for 5-15 s with native stereo audio.
• Audio and video references are honored by MiniMax H3 and H3 Max; image, video and audio appear in supported_input_references for H3 Max.
• Every Sume model bills provider list times 1.25.

A pre-flight check

Reference limits fail late if you let the provider reject them, so check before you submit. This script validates a reference set against the vendor caps above. It only uses the standard library.

import os

CAPS = {"image": 30, "video": 50, "audio": 15}  # MB, from the vendor guide
COUNTS = {"image": 9, "video": 3, "audio": 3}

def check(paths):
    problems = []
    seen = {"image": 0, "video": 0, "audio": 0}
    for kind, path in paths:
        seen[kind] += 1
        mb = os.path.getsize(path) / 1_000_000
        if mb > CAPS[kind]:
            problems.append(f"{path}: {mb:.1f} MB over {CAPS[kind]} MB")
    for kind, n in seen.items():
        if n > COUNTS[kind]:
            problems.append(f"{n} {kind} references, max {COUNTS[kind]}")
    return problems

if __name__ == "__main__":
    print(check([]) or "ok")

How to use the numbers

Reference-heavy prompts are where per-model limits bite. Nine images, three clips and three audio files is a lot of input, and each extra reference adds a chance that one URL is unreachable or one file is over its size cap. Start with the references that carry the most information, usually one product image and one style frame, and add more only when the output drifts.

Also remember the prompt cap. A 7,000 character prompt is long enough to contain a full shot list, but a model follows a short, specific prompt more reliably than a long, vague one. Put camera, subject and lighting first, and keep the rest for continuity details.

When the numbers disagree

If the vendor page and the Sume catalog differ, the catalog is what your request is validated against. Read GET /v1/videos/models for the live supported_durations and supported_input_references before you hard-code a limit. The Sume docs state their own reference fields on the legacy Video 1.0 route as 1-9 images, 1-3 videos and 1-3 audio URLs, with audio allowed only alongside at least one image or video reference.