MiniMax H3 reference limits: 9 images, 3 videos, 3 audios on Sume
MiniMax H3 reference-to-video on Sume takes up to 9 images, 3 videos and 3 audio clips, 12 in total. The duration rules, the extra-image fee and a request.

MiniMax H3 (minimax-h3) on Sume accepts up to 9 reference images, up to 3 reference videos and up to 3 reference audio clips in one request, with a combined cap of 12 references, and audio cannot be the only reference. Each reference video or audio clip must be 2 to 15 seconds long, and the videos together, and the audios together, can total at most 15 seconds. These constraints are listed on the minimax-h3 row of the Video Router catalog (GET /v1/video-router/models); the Video Router docs cover the model's 5 to 15 second range and native 480p/768p resolutions (read 2026-10-06).
MiniMax H3 launched on 2026-07-31 and, per Magic Hour's tracker (read 2026-10-06), makes 15-second clips at 2K with native stereo audio; the open weights followed on 2026-08-03. On Sume the hosted model is native 480p or 768p for 5 to 15 seconds, with 2K and 4K as priced upscales.
The limits in one table
These are the caps the catalog states for reference-to-video. A request over the 12-reference total is rejected with unsupported_capability before a job is created.
| Input | Maximum | Per-item rule | Notes |
|---|---|---|---|
image_url | 9 | none stated | first 5 images free, extra ones priced |
video_url | 3 | 2-15 s each, combined 15 s | counts toward the 12 total |
audio_url | 3 | 2-15 s each, combined 15 s | cannot be the only reference |
| All references together | 12 | images + videos + audios | one request |
What the extra-image fee is
The provider includes the first five reference images in a reference-to-video job and charges a list fee of $0.08 for each image after the fifth. Sume applies its usual 1.25 multiple to provider list prices, so check the catalog for the billed amount before a job that uses more than five. A request with 9 images therefore carries four extra-image fees on top of the per-second rate. If your look can be anchored with five images, stay at five.
Per second, the base rate is the same as for plain text-to-video: $0.05 list at 480p and $0.06 at 768p, which Sume bills at $0.0625 and $0.075. A 10-second 768p reference job is $0.75 before any extra-image fees.
A reference-to-video request
References go in input_references as typed entries. This sends two images and one audio clip, which is within every cap above.
import os, requests
r = requests.post(
"https://api.sume.com/v1/videos",
headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
json={
"model": "minimax-h3",
"prompt": "The two friends from the references laugh at a cafe table",
"duration": 8,
"resolution": "768p",
"aspect_ratio": "16:9",
"input_references": [
{"type": "image_url", "image_url": {"url": "https://example.com/a.jpg"}},
{"type": "image_url", "image_url": {"url": "https://example.com/b.jpg"}},
{"type": "audio_url", "audio_url": {"url": "https://example.com/line.mp3"}},
],
},
timeout=60,
)
print(r.status_code, r.json())Budgeting a reference set
Count references before you count dollars. Nine images, three videos and three audios would be fifteen files, so you can never fill every slot; the 12-file total is the binding limit. A practical split for a character-driven clip is four or five images for the subject and setting, one short reference video for motion, and one audio clip for the voice or rhythm. That stays within the five free images, so the only price driver is the clip length.
The per-second math is simple enough to do before you submit. At 768p the billed rate is $0.075 a second, so a 15-second job is $1.125. At 480p it is $0.0625 a second, or $0.9375 for 15 seconds. Add $0.10 for each image beyond the fifth ($0.08 list times the 1.25 margin). A 15-second 768p job with nine images is therefore $1.125 plus four times $0.10, or $1.525. Both figures use the billed amount, and the catalog's billable_formula field is the final word if the rate card changes.
Why the audio rule matters
Audio can steer a clip, but it cannot be the only reference. A request with just an audio_url is rejected, so pair it with at least one image or video. That also means H3's audio reference is a way to condition the soundtrack and timing of a clip that already has a visual anchor, not a route to audio-only generation.
A reference video is the more demanding input. It must be at least 2 seconds, and the three together cannot exceed 15. If you have a longer source clip, trim it first. Sume has a video trim tool for exactly this.
H3 or H3 Max for references
MiniMax H3 Max (minimax-h3-max) has the same reference envelope and adds a 1080p option, at $0.16 list a second. H3 stops at 768p natively. If your references are detailed, such as a product label or a face, the larger frame may be worth the higher rate; if they set mood and color, 768p is enough. The question of whether the weights run locally is separate, and MiniMax H3 open weights covers it. To check any value before you send it, use the catalog.
Sources
Related posts
More in Models
- Native 4K AI video API: LTX-2.5 claims 4K, what Sume offers instead
LTX-2.5 is listed up to 4K but is not in Sume's catalog. Sume's 4K options are Gemini Omni Flash 1.1 and MiniMax H3 4K upscales. Limits and per-second prices.
- New AI video models since February 2026: dates and which Sume lists
Kling 3.0, Seedance 2.5, MiniMax H3 and LTX-2.5 launched in 2026. Release dates, length limits and which of them Sume's video catalog lists today.
- Retired AI model id to Sume id: GPT Image, Gemini and Sora lookup
One lookup table that maps retired gpt-image-1, gemini-2.5-flash-image and sora-2 ids to Sume ids, with dates, so a failing call tells you what to switch to.
- seedream-v4 or bytedance-seed/seedream-4: which Sume id to store
Sume accepts legacy bare ids like seedream-v4 and seedream-4-5-edit, but the catalog lists org/slug ids. Store the canonical id; Python normalizer inside.
Written by Sume