MiniMax H3
MiniMax H3 is MiniMax’s Hailuo video generation model for clips of 5 to 15 seconds at 768P or 2K, where the soundtrack is composed together with the picture rather than laid over it. Use it when the audio is the point of the shot: ambience, effects or a voice that has to sit with the motion, and 2K when the picture has to hold up large.
It runs in the editor on the left at 16 credits a second for 768P and 23 for 2K, so a 5-second 768P clip is 80 credits and a 15-second 2K one is 345. Reference video is billed by its own length at 13 or 22 credits a second, and the sixth and later reference images cost 9 credits each; the first five are free.
Sound first
Most models here that produce audio generate it as a by-product of the picture. H3 is the one built around the pairing: the vendor describes picture and sound as composed in a single pass, and in our test the water and ambience track sat with the motion in a way that reads as designed rather than dubbed. It also takes reference audio as an input, up to three clips, alongside up to nine images and three videos.
Frame shapes run from 21:9 through 16:9, 4:3, 1:1, 3:4 and 9:16, plus adaptive when the shape should follow the input. 2K is the largest output in the studio; nothing else here goes past 1080p.
What it costs
Per second of output, plus reference material where it is billed. The first five reference images are free.
| Item | 768P | 2K |
|---|---|---|
| Output, per second | 16 credits | 23 credits |
| 5 seconds | 80 | 115 |
| 15 seconds | 240 | 345 |
| Reference video, per second of its length | 13 credits | 22 credits |
| Reference images 6 to 9, each | 9 credits | 9 credits |
Our test, including the surprise
The hero clip is a text-to-video job: one sentence about a paper boat, 768P, 5 seconds requested, 16:9. It cost 80 credits and took about seven minutes, the longest wait in the studio. It came back as a 1376×768 MP4 with composed audio — and 10.1 seconds long, double what we asked for. The vendor billed it as 5 seconds and so did we, so the extra length was free, but do not plan on the returned length matching the request; the page records what the file actually is.
The same sentence went through Grok Video 1.5 and Seedance 2.0, so the three text models can be compared on one prompt.
A second job on 2026-09-24 added a 5-second reference clip: 80 credits for the 5-second output plus 65 for the clip’s own length at 13 a second, 145 in all — and again a 10.1-second clip came back for a 5-second request.
| Field | Value |
|---|---|
| Input | Prompt only, 16:9 |
| Settings | 768P, 5 seconds |
| Cost | 80 credits ($0.80) |
| Wall time | About 7 minutes (vendor reported 387 s) |
| Output | 1376×768, 10.1 s, MP4 with composed audio |
MiniMax H3 or Seedance?
Seedance 2.0 is cheaper at the small size (75 against 80 for 5 seconds) and offers 1080p; Seedance 2.5 goes to 30 seconds. H3 is the choice when the sound has to be right, when 2K matters more than 1080p, or when a 21:9 frame is the brief. For a quick test of any of these, Grok Video 1.5 at 25 credits comes back in a minute rather than seven.
How to run MiniMax H3 here
- 1. Write the scene with its sound in it Name what should be heard as well as seen; the model composes both.
- 2. Add reference images if the subject is fixed Optional, up to nine; five are free.
- 3. Set 768P and 5 seconds 80 credits. Go to 2K for the take you keep.
- 4. Create and come back Five to eight minutes. Leave the page; the clip waits in History.
When not to use this
- Anything you need back quickly. Seven minutes a clip is the slowest turnaround here; draft on Grok Video 1.5 in a minute.
- An exact clip length. Our 5-second request returned 10 seconds. If the edit needs a precise duration, trim afterwards or use another model.
- 1080p specifically. H3 offers 768P and 2K; Seedance 2.0 and Wan 3.0 offer 1080p.
- Cheap reference video. Each second of footage you attach is billed at 13 or 22 credits on top of the output.
- Footage of a real person who has not agreed to it.
Questions people ask
How much does MiniMax H3 cost?
16 credits a second at 768P and 23 at 2K, where 1 credit is one US cent. A 5-second 768P clip is 80 credits, a 15-second 2K clip is 345. Reference video adds 13 or 22 credits per second of its length, and reference images beyond the fifth add 9 each. Failed jobs are refunded.
What is 768P?
An output of 768 pixels on the short side: our 16:9 test came back at 1376×768. 2K is the larger option and the largest output in the studio.
Does MiniMax H3 generate sound?
Yes, composed with the picture in one pass, which is the model’s distinguishing feature. It also accepts reference audio as an input.
How long does it take?
About seven minutes for a 5-second 768P clip on our test, the longest in the studio. You can leave the page; the result waits in History.
Why did my clip come back longer than I asked?
It happens: our 5-second request returned 10.1 seconds. You are billed for the length you asked for, not the length returned. Trim in an editor if the cut needs to be exact.
Related tools
Checked 2026-09-23.