MiniMax H3
MiniMax H3 makes 480p or 720p clips with sound, from 2 to 15 seconds, in 16:9 or 9:16. It accepts a text prompt, an opening image, an opening and closing image, or up to eight reference pictures. With enhance_prompt on (the default), a short prompt is expanded into timed shots with camera moves and sound. The model runs on AI HUB's own servers and is not offered in the United States, the European Union, the United Kingdom or South Korea.
Model id: minimax-h3
Parameters
Only model and prompt are required. Omitted optional params use the default. An unsupported value returns 400. These values are the same ones GET /v1/models?type=video reports.
| Parameter | Allowed | Default | Required |
|---|---|---|---|
resolution |
480p, 720p | 480p | No |
duration |
2-15 | 5 | No |
audio |
true | true | No |
enhance_prompt |
true, false | true | No |
seed |
0-2147483647 | — | No |
ratio |
16:9, 9:16, 1:1 | 16:9 | No |
Some combinations are constrained. See the overview for details.
Constraints: 720p ⇒ aspect ["16:9","9:16"], ratio ["16:9","9:16"]
Attachment roles and upload limits
| Role | Content block | Max per request | Supported | Required |
|---|---|---|---|---|
Opening
first_frame
|
image_url |
— | Yes | Optional |
Ending
last_frame
|
image_url |
— | Yes | Optional |
Reference image
reference_image
|
image_url |
8 | Yes | Optional |
Reference video
reference_video
|
video_url |
— | No | Optional |
Source video (edit/motion reference)
source_video
|
video_url |
— | No | Optional |
| Combination | Allowed | Notes |
|---|---|---|
| Opening + Ending | Yes | Sets the clip's opening and closing image. This is only offered on models that accept a last frame. Other models refuse the second image. |
| Opening + Reference image | Yes | Refused. A picture used as a frame and a picture used as a likeness reference are different modes, and no model accepts both at once. |
| Opening + Reference video | No | Allowed. A video reference is a different medium from a frame, so the two do not compete. |
| Reference image | Yes | Up to the per-request maximum in the table above. The cap differs per model, so one model's number never applies to another. |
| Source video (edit/motion reference) | No | The source clip's own duration is billed on top of the output, so a 5s edit of a 5s clip bills as 10 seconds. |
| Source image (photo to edit) | No | Never accepted on a video model, whatever its other capabilities. Editing a photo belongs on the image endpoint. |
The role token distinguishes two attachments. First frame and last frame both travel as an image_url content block, so the block alone cannot identify them.
On this endpoint, a request may carry only one image_url attachment, and a second one is rejected even though /media allows more. This is the most common cause of support requests.
Sending a role a model does not support returns 422 model_capability_unsupported.
Per-file and total byte ceilings differ per model. One model's numbers never apply to another.
| Field | Value |
|---|---|
| Accepted image types | image/jpeg, image/png, image/webp |
| Max per image | 30 MB |
| Max total across all images | 64 MB |
| Dimensions | 300–6000 px |
| Aspect ratio (width / height) | 0.4–2.5 |
Example
curl https://console.aimodelhub.ai/v1/videos \
-H "Authorization: Bearer $AIHUB_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"minimax-h3","prompt":"a drone shot over a canyon"}'Metered once at this model's list price on success and debited from the wallet. See the Video overview for the full request/response shape.