Skip to main content

MiniMax H3 AI Video Generator

Type a prompt or start from a photo. A 5-second clip costs 10 credits at 720p and 15 at 1080p, and the sound comes out of the same render.

Released:
31 July 2026, by MiniMax
Output:
720p or 1080p, 5 to 15 seconds
Inputs:
Text, a first-frame photo, or reference photos

From 10 credits for a 5-second clip. Exact price shown before you generate.

What MiniMax H3 made in our tests

We rendered these on 16 October 2026: five clips from text and one from a photo, each 5 seconds at 720p. Each prompt ran once and every result is shown. Press play for sound, or Try this prompt to load it into the box above.

A dog in the surf at sunset

16:9, 720p, 5 s, text to video, 44 s to render

Backlit spray and fur hold up well. We asked for a puppy and got a grown retriever.

Prompt we used

A golden retriever puppy runs through shallow ocean waves at sunset, water splashing in slow motion, the camera tracks low beside it, the sound of waves and happy barking.

Noodles tossed in a wok, vertical

9:16, 720p, 5 s, text to video, 43 s to render

The toss, the steam and the busy background all read as a real stall. The flame stays under the wok; there are no sparks.

Prompt we used

Close-up of a street cook tossing noodles in a flaming wok at a night market, steam and sparks rising, warm string lights behind, loud sizzling and crowd chatter.

A spoken line at a rainy bus stop

16:9, 720p, 5 s, text to video, 43 s to render

She turns, walks toward the lens and speaks with matching lip movement. A speech-to-text pass heard "Is this the last bus, bus tonight?", so one word doubled.

Prompt we used

A young woman in a yellow raincoat waits at a rainy bus stop at night, she looks into the camera and says: "Is this the last bus tonight?" Rain drumming on the shelter roof, headlights passing.

A product macro on wet black glass

1:1, 720p, 5 s, text to video, 43 s to render

Clean, label-free product look with crisp droplets. The turn is very slight: you mostly see it in the tube inside the bottle.

Prompt we used

A plain glass perfume bottle with no label turns slowly on a wet black surface, water droplets on the glass, soft rim light, macro lens, quiet ambient hum.

A skateboard trick at golden hour

9:16, 720p, 5 s, text to video, 43 s to render

Body motion and the low follow-cam are convincing. The board pops and lands under his feet, but it reads more as a hop than a clean kickflip.

Prompt we used

A skateboarder in a grey hoodie lands a kickflip in an empty concrete skatepark at golden hour, the camera follows at board height, wheels clatter and roll.

Photo to video: a bicycle on a sunny street

9:16, 720p, 5 s, from one photo, 46 s to render

The first frame is a still we made with GPT Image 2.5. The bike, lemons and blue door stay put while the camera moves in and the cat walks through.

Prompt we used

The camera slowly pushes in toward the bicycle while a light breeze moves the plants, a grey cat walks past the front wheel, birds chirp.

MiniMax H3 price per clip, in credits

Credits per clip, read from the same rate card the Generate button uses. Reference photos keep a person or product the same across clips and cost more.

ModeResolution5 s8 s10 s15 s
Text or photo to video720p10162030
Text or photo to video1080p15243045
With reference photos720p406480120
With reference photos1080p6096120180

Credits come from a plan or a top-up pack. Every model and price is on the models page.

MiniMax H3 at 720p versus 1080p

720p costs 10 credits for 5 seconds and 1080p costs 15. We ran the dog prompt at both. The 1080p file holds finer fur and droplets, worth it on a large screen; on a phone feed 720p is hard to tell apart. 1080p also took 85 seconds to render against 44.

Close crop of the dog's head from the 720p file720p file
Close crop of the dog's head from the 1080p file1080p file
The same share of the frame from each file, shown at the same size. Two separate runs of one prompt, so the pose differs; look at the fur and the spray.

MiniMax H3 specs on CastTake

SpecMiniMax H3
MakerMiniMax, released 31 July 2026
Modes on CastTakeText to video, photo to video (first frame), reference photos
Resolution720p or 1080p
Length5, 8, 10, 15 seconds
Shapes16:9, 9:16 and 1:1
SoundGenerated in the same render: ambience, effects and short spoken lines
Render time in our run43 to 46 s at 720p, 85 s at 1080p (5-second clips)

MiniMax H3 versus Seedance 2.5 and H3 Max

Seedance 2.5 is the other new video model on our shelf, and H3 Max is the variant people ask about. Prices are from our rate card; the H3 Max column is what MiniMax publishes, since we do not offer it.

Compared onMiniMax H3Seedance 2.5
Price, 5 s at 720p10 credits325 credits
Resolution on CastTake720p and 1080p720p
Length5, 8, 10 or 15 s5, 8, 10 or 15 s
Photo inputFirst frame or reference photosFirst frame or reference photos
SoundYes, tested on this pageListed for dialogue; we have not run our own test yet
Pick it forMany tries, single shots, product clipsMulti-shot scenes where one render must tell a story
Compared onMiniMax H3MiniMax H3 Max
What it isThe base modelA faster variant post-trained on top of H3
Top resolution1080p on CastTake768p, per MiniMax
SpeedAbout 45 s for 5 s at 720p in our runSold as the faster of the two
On CastTakeYes, in the studioNot offered yet

MiniMax H3 prompt tips

Every prompt above follows the same five parts. Copy any sample prompt and swap the words in each part.

PartExample from our testsWhy it matters
Subject and actionA skateboarder in a grey hoodie lands a trickOne subject doing one thing keeps five seconds readable.
Camerathe camera follows at board heightH3 followed every camera note we gave it: low tracking, push-in, close-up.
Place and lightempty concrete skatepark at golden hourLight words changed the look more than style words did.
Soundwheels clatter and rollSound is made in the same render, so name the noises you want.
Spoken lineshe says: "Is this the last bus tonight?"Quote short lines exactly. Keep them under about ten words for a 5-second clip.

How to use it

  1. Write your prompt in the box at the top of this page, or press Try this prompt under any sample.
  2. Press Open in studio. MiniMax H3 is already selected there, with your prompt filled in.
  3. Pick 720p or 1080p, the length and the shape. Add a photo if you want it to be the first frame.
  4. Check the credit price on the Generate button, press it, and the clip appears in your results in about a minute.

Limits we hit with MiniMax H3

What went sideways, so you know what to check before you post a clip.

  • It sometimes simplifies what you ask for: "puppy" became a grown dog, "kickflip" became a hop, "turns slowly" became a slight turn.
  • Spoken lines can stutter. Our one dialogue test doubled a word; listen before you publish.
  • Render time doubles at 1080p: 85 s against about 45 s at 720p for the same 5-second prompt.
  • The same prompt gives a different shot each run, so a 1080p re-render will not match an earlier 720p take frame for frame.
ClipInputShape and sizeRender time
Dog in the surfText16:9, 720p, 5 s44 s
Wok tossText9:16, 720p, 5 s43 s
Bus stop lineText16:9, 720p, 5 s43 s
Perfume macroText1:1, 720p, 5 s43 s
Skate trickText9:16, 720p, 5 s43 s
Bicycle streetPhoto9:16, 720p, 5 s46 s
Dog in the surfText16:9, 1080p, 5 s85 s

MiniMax H3 questions

What is MiniMax H3?

MiniMax H3 is the video model MiniMax released on 31 July 2026. It turns a text prompt or a still photo into a short clip and generates the sound in the same pass, including short spoken lines. On CastTake you pick it in the studio, choose 720p or 1080p and a length of 5, 8, 10 or 15 seconds, and see the credit price before you press Generate. We rendered six clips with it on 16 October 2026, plus a 1080p re-run of one prompt, and put every result on this page with its prompt, so you can judge the model from real output rather than a feature list.

Is MiniMax H3 free?

Not on CastTake, and we would rather say so up front. Sign-in credits do not cover video renders: your first credit pack opens every video model, and pack credits never expire. A 5-second 720p clip costs 10 credits, the lowest price of any video model we sell, so a pack goes a long way while you test prompts. Before you buy, watch the six clips above and copy a prompt you like, so your first paid render starts from something that already worked. The studio shows the exact price on the Generate button, and nothing is charged until you press it.

How much does MiniMax H3 cost per video?

A clip from text or one photo costs 10 credits for 5 seconds at 720p and 15 at 1080p. Longer clips scale with length, so 15 seconds at 720p is 30 credits. Using reference photos to keep a person or product consistent costs more: 40 credits for 5 seconds at 720p and 60 at 1080p. The table above lists every length, and the studio shows the exact total on the Generate button before anything renders. Credits come from a plan or a top-up pack on the pricing page.

Does MiniMax H3 make sound and speech?

Yes. Every clip on this page came back with an audio track made in the same render, with no separate step. Ambience and effects matched the scene: surf, a sizzling wok, rain on a shelter roof. In the bus stop test we quoted a short line and the woman spoke it with matching lip movement, though our transcript heard one word twice. Name the sounds you want in the prompt, put dialogue in quotation marks, and keep lines short enough to say comfortably inside the clip length you picked.

Is there a MiniMax H3 API?

CastTake does not sell API access, only the studio, where you type a prompt, pick the model and download the finished MP4. MiniMax offers its own API for developers who want to build H3 into an app, and its documentation covers keys, request limits and pricing. If you only need clips rather than an integration, the studio is simpler: no code, no key to manage, the price shown before each render, and your finished videos kept in your account history so you can download them again later.

MiniMax H3 or Seedance 2.5: which should I use?

Start with H3 unless you need a multi-shot scene from one render. On CastTake a 5-second 720p clip is 10 credits on H3 and 325 on Seedance 2.5, so you can try dozens of H3 prompts for the price of one Seedance clip. H3 also renders 1080p, while Seedance 2.5 is sold at 720p here. Seedance 2.5 is described as stronger at multi-shot storytelling with dialogue. A practical order is to find the idea on H3, then spend on Seedance 2.5 only when the story truly needs several shots inside a single clip.

More from CastTake