MiniMax H3 is MiniMax's general-purpose multimodal video-generation model. It can work from text, images, video, and audio references, then create short videos with native sound. Through HiAPI, the model ID is minimax-h3, and it uses the same asynchronous task interface as other image, video, and audio models.
For a new clip from a written idea, send only a prompt. When product appearance, brand treatment, or a specific shot transition matters, use first-and-last-frame control or multimodal references. The current model supports 4 to 15 second output, native 2K, and aspect ratios including 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16. Try the product-shot prompt below: Generate with this prompt → Select 5 seconds, 2K, and 16:9. Read the API reference before integrating.
What MiniMax H3 is useful for
H3 is useful when a short video needs coordinated control over the picture, motion, and sound. Typical examples include a product demo with a deliberate camera move, a brand visual that carries through an established look, or a new shot guided by a subject image, motion reference, and audio reference.
For e-commerce, advertising, and product teams, the goal is not simply to animate an image. State what must stay fixed, such as product shape, logo position, subject identity, camera angle, and lighting. Describe changes separately, such as motion, background, camera move, or dialogue. That gives a reviewer a clear way to identify what drifted.
One model, three common workflows
| Workflow | Good for | Key input |
|---|---|---|
| Text to video | Creating a new scene from a script or creative direction | prompt |
| First and last frame | Controlling a product's start and end state or a scene transition | first_frame_image, optional last_frame_image |
| Multimodal reference video | Combining product, subject, motion, or audio references | image_urls, video_urls, audio_urls |
First-and-last-frame control and multimodal reference are separate modes and cannot be combined in one request. A last frame requires a first frame. Multimodal reference accepts up to 5 images, 3 video clips, and 3 audio clips. Audio cannot be the only reference; it needs an image or video reference as well. Each reference video can be 2 to 15 seconds, with a 15-second total limit across videos. Reference audio has a separate 15-second total limit.
Start with a minimal text-to-video task
This example creates a five-second 16:9 product shot. prompt is required in every mode and accepts up to 7,000 characters. A successful create response (code: 200) only means the service accepted the task. Check that data.taskId is a non-empty string before following its status. If the POST response is lost, check your task records before submitting again.
Prompt:
A matte navy coffee mug rests on a pale gray stone table. The camera slowly pushes in from a medium shot. Morning light enters from the left. Keep the mug still, soften the background, no text.
Generate with this prompt → Select 5 seconds, 2K, and 16:9.
curl --fail-with-body --connect-timeout 10 --max-time 60 -X POST "https://api.hiapi.ai/v1/tasks" \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax-h3",
"input": {
"prompt": "A matte navy coffee mug rests on a pale gray stone table. The camera slowly pushes in from a medium shot. Morning light enters from the left. Keep the mug still, soften the background, no text.",
"duration": 5,
"resolution": "2K",
"aspect_ratio": "16:9",
"watermark": false
}
}'
For local testing, query GET /v1/tasks/{taskId}. In production, set a top-level callback.url and make your handler idempotent by taskId. After a task reaches a terminal state, inspect status: download the URL of the type: "video" item in data.output after success; use the returned error to correct the request after fail instead of blindly replaying the same invalid input.
How to use reference material well
First and last frames suit a shot that needs to move from state A to state B. For example, a launch clip can start with an unlit lamp and end with warm light across a desk. The first frame anchors the beginning, the last frame anchors the destination, and the prompt describes movement, camera work, and light between them.
Multimodal references work better for brand or subject consistency. Put the product image in image_urls, a camera-rhythm reference in video_urls, and a sound or mood reference in audio_urls. State the role of each asset: “Keep the bag shape and metal clasp from the reference image, follow the sideways push-in rhythm of the reference video, and create a quiet product film.”
Do not add every possible reference at once. Each additional asset adds constraints to reconcile. First verify subject preservation with one image, then add motion or audio references so it stays easy to locate failures.
What a useful prompt includes
A reusable video prompt usually covers four things: subject, action, camera, and sound or atmosphere. The subject section states what cannot change. The action describes the event. The camera section specifies framing and movement. The sound section covers dialogue, ambience, or music direction.
For a product ad, write: “Keep the white earbud case, orange shell, and NOMA mark from the reference image unchanged. The camera slowly pushes in from a medium table shot. The case opens and both earbuds remain in place. Natural window light, quiet room ambience, no voiceover, no extra text.” This gives a clearer acceptance target than “make a beautiful earbud ad.”
Before putting H3 into production
Run a small acceptance set: one text-to-video task, one first-and-last-frame task, and one multimodal reference task. Record input assets, output duration, status changes, failure messages, output URL, and final invoice. Only expand to batch generation after the modes your product needs have passed this test.
Pricing is based on output seconds and the current model tier, so do not hard-code a rate in this article. Review the live MiniMax H3 model page before submission and use task records and account billing as the production source of truth. For asynchronous tasks, callbacks, and failure handling, read the Unified Async API guide.
Frequently asked questions
What inputs does MiniMax H3 support?
It can generate from text alone or use first frames, first-and-last frames, and reference images, videos, or audio. First-and-last-frame fields cannot be mixed with multimodal reference fields.
Can MiniMax H3 generate sound?
MiniMax H3 supports video with native sound and can use audio material in multimodal-reference mode. Audio references need an accompanying image or video reference.
How long can one video be?
Current output duration is 4 to 15 seconds. Longer sequences usually work better as individual shots generated separately, then edited together in your own workflow.
Should I poll or use a callback?
Polling works for local development and low-frequency tasks. In production, use callback.url for terminal notifications and retain task-detail queries as a recovery path when callbacks fail.
Generate with this prompt → Select 5 seconds, 2K, and 16:9; read the API reference, or get an API key.
