MiniMax H3 reference inputs: 12 files combined, not 9 + 3 + 3
MiniMax H3 allows 9 images, 3 videos and 3 audio clips as references, but a 12-file combined cap applies. How to stay inside it, and how Sume counts.
MiniMax's H3 guide caps reference inputs by type (up to 9 images, 3 videos, 3 audio clips) and also at 12 reference files in total (read 2026-10-02). Add the per-type caps and you get 15, so a request that maxes every type is over the combined limit. Sume's Video 1.0 docs list 1 to 9 image, 1 to 3 video and 1 to 3 audio reference URLs, and I found no combined cap documented there, so keep your own total at 12 or fewer.
What are all of MiniMax's reference limits?
The numbers below come from the same guide. Video and audio references are also bounded by clip length, which is easy to miss.
| Input | Count | Per-file limit | Other rule |
|---|---|---|---|
| Reference images | Up to 9 | 30 MB | None listed |
| Reference videos | Up to 3 | 50 MB, 2 to 15 s | Total 15 s or less |
| Reference audio | Up to 3 | 15 MB, 2 to 15 s | Total 15 s or less |
| All references | 12 files | Request body 64 MB | Combined cap |
How does that price out?
References are not free of cost on every route. Sume's router notes list an extra $0.08 per image after the fifth for minimax-h3 reference-to-video, taken from the provider list and then billed with Sume's 1.25 multiplier (Sume docs, fal list as of 2026-08-25). MiniMax's own pay-as-you-go page lists H3 at $0.08 per second at 768P and $0.13 at 2K, and the lines I read give per-second prices only (read 2026-10-02). Confirm the current surcharge in GET /v1/video-router/models before budgeting a reference-heavy batch.
What does Sume accept?
Per the Video 1.0 page, reference_image_urls takes 1 to 9 URLs, reference_video_urls 1 to 3, and reference_audio_urls 1 to 3 but only alongside at least one image or video. All must be public HTTPS URLs. Sume's video generation docs say audio and video references are honored by the Seedance 2.x models, Wan 3.0, MiniMax H3 and MiniMax H3 Max.
Not every model takes references: kling-3 does not, and Gemini Omni Flash 1.1 takes video but not audio. Read supported_input_references per model rather than assuming.
A safe working set
- Pick the three or four references that carry the identity or style; more rarely helps.
- Keep video and audio clips under 15 seconds in total each.
- Count files before you send: 6 images, 3 videos and 3 audio clips is already 12.
- If a request is refused, cut the least important reference first and resend with the same idempotency key.
Sources
Related posts
More in Models
- MiniMax H3 sound design prompts: direct the audio like the picture
fal's H3 guide says to direct audio as deliberately as picture: name sonic elements, not 'music'. What it looks like in a Sume request, and what you can't set.
- MiniMax-Hailuo-2.3-Fast: an image-to-video model id, no Sume twin
MiniMax lists MiniMax-Hailuo-2.3-Fast in its image-to-video API but not in its text-to-video enum. What that means, and what Sume offers for first frames.
- MiniMax Hailuo 2.3 vs H3 API: model ids, durations, prompt limits
MiniMax's API still lists MiniMax-Hailuo-2.3 (6 or 10 s, 2,000-character prompt) beside MiniMax-H3 (4-15 s, 7,000). Which one Sume lists, and why it matters.
- Hailuo camera commands: the 15 bracket options like [Push in]
MiniMax's Hailuo API reads 15 camera commands in square brackets, such as [Truck left] or [Push in]. The full list, and what to do on Sume.
Written by Sume