MiniMax-H3 Video Generation

MiniMax-H3 (MiniMax-H3) is MiniMax's Hailuo-series video generation model. It supports text-to-video, image-to-video (first frame) and video reference, outputting 768P by default with an optional 2K mode. It uses the unified POST /v1/videos endpoint and returns an OpenAI Video style response (id / status / metadata.url), sharing the same task polling mechanism as other video models.

Billing

Billed by actual generated duration (seconds), with different unit prices per resolution tier (768P / 2K). Exact unit prices and billing expressions are shown on the corresponding model card in the console "Model Plaza" (prices are intentionally omitted from this documentation).

Endpoints

• Create task: POST $URL/v1/videos
• Query task: GET $URL/v1/videos/{task_id}

Compatible alias

An OpenAI Sora-style alias is also supported: POST $URL/v1/video/generations, GET $URL/v1/video/generations/{task_id} — equivalent to the above.

Supported Models

Model name	Positioning	Resolutions
MiniMax-H3	Video generation, 768P by default, 2K optional	768P / 2K

The resolution tier is selected via metadata.resolution (see "Parameters" below).

Parameters

Parameter	Type	Required	Description
model	string	yes	MiniMax-H3
prompt	string	yes	Video description; recommend specifying subject, action, scene, and camera movement
image / images	string / array	no	Reference image URL(s), used as the first frame for image-to-video
input_reference	string	no	Reference video URL, for video reference generation
seconds	string	no	Output duration in seconds, e.g. "5", "10"; defaults to 5 seconds
metadata.resolution	string	no	Resolution tier: 768P (default) / 2K
metadata.ratio	string	no	Aspect ratio, defaults to 16:9, e.g. 9:16, 1:1
metadata.duration	number	no	Override output duration (seconds)
metadata.aigc_watermark	boolean	no	Whether to add an AIGC content watermark
metadata.callback_url	string	no	Callback URL for task completion (upstream push)

Reference material

Publicly accessible image/video URLs are supported (the upstream server fetches them; intranet or local addresses will fail).

Request Examples

Text-to-video

curl

# 1. Create video task (768P, 5 seconds)
curl $URL/v1/videos \
  -H "Authorization: Bearer $YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-H3",
    "prompt": "A dolphin leaping out of the sea at sunset, golden splashes, slow motion, cinematic",
    "seconds": "5"
  }'

# 2. Query task status (replace {task_id} with the id returned above)
curl $URL/v1/videos/{task_id} \
  -H "Authorization: Bearer $YOUR_API_KEY"

Python

import time
import requests

url = "$URL/v1/videos"
headers = {
    "Authorization": "Bearer YOUR_API_KEY",
    "Content-Type": "application/json",
}
payload = {
    "model": "MiniMax-H3",
    "prompt": "A dolphin leaping out of the sea at sunset, golden splashes, slow motion, cinematic",
    "seconds": "5",
}

task = requests.post(url, json=payload, headers=headers).json()
task_id = task["id"]

while True:
    result = requests.get(f"{url}/{task_id}", headers=headers).json()
    if result["status"] in ("completed", "failed"):
        break
    time.sleep(8)

print(result)

Image-to-video (first frame)

curl $URL/v1/videos \
  -H "Authorization: Bearer $YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-H3",
    "prompt": "Make the robot in the picture stand up and wave at the camera",
    "image": "https://example.com/ref.png",
    "seconds": "5"
  }'

Video reference

curl $URL/v1/videos \
  -H "Authorization: Bearer $YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-H3",
    "prompt": "Transform the scene in the video into a neon night lighting effect",
    "input_reference": "https://example.com/clip.mp4",
    "seconds": "5"
  }'

Switching to 2K

Use metadata.resolution to select the resolution tier:

curl $URL/v1/videos \
  -H "Authorization: Bearer $YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-H3",
    "prompt": "Aerial drone shot of snow-capped mountains above the clouds, golden sunrise, breathtaking panorama",
    "seconds": "5",
    "metadata": {
      "resolution": "2K"
    }
  }'

Response Fields

Create and query responses share the same structure (OpenAI Video style):

Field	Type	Description
id	string	Task ID
status	string	Task status (see below)
metadata.url	string	Video URL after completion (only on completed)
progress	integer	Generation progress, 0–100
completed_at	integer	Completion timestamp (seconds, only on completed)
error.message	string	Failure reason (only on failed)

Task Status

Status	Meaning	Suggestion
queued	Queued	Keep polling
in_progress	Generating	Keep polling
completed	Success	Get the video URL from metadata.url and download it as soon as possible
failed	Failed	Check error.message

Poll every 5–10 seconds

Video URLs are time-limited — download and store the result as soon as possible after completion.

Response Example

{
  "id": "task_xxxxxxxx",
  "status": "in_progress",
  "progress": 50
}

After completion:

{
  "id": "task_xxxxxxxx",
  "status": "completed",
  "progress": 100,
  "metadata": {
    "url": "https://cdn.example.com/.../result.mp4"
  },
  "completed_at": 1786431715
}

FAQ

Q: What reference materials are supported?

Reference images (first-frame image-to-video) and reference videos (video reference generation) — just pass publicly accessible URLs (see image / images / input_reference above).

Q: How long does a task take?

It depends on duration and resolution. A 768P 5-second task usually completes within 1–3 minutes; 2K takes longer. Poll the query endpoint for the final status.

Q: Do generated video links expire?

Yes. Video URLs are time-limited, so download and store the result as soon as possible.

Q: Is there an asset library?

No standalone asset library — use publicly accessible image/video URLs as reference material.