MiniMax Video Generation

POST /v2/video_generation — MiniMax-H3 native video protocol (text / image / reference to video)

MiniMax-H3 generates video from text, images (first frame / last frame / first+last), or multimodal references. This family uses its own native entrance — the request shape differs from the /api/v1 families.

Models

Model	Notes
`MiniMax-H3`	MiniMax H3 (model name is case-insensitive; `minimax-h3` works too)

Create a job

POST/v2/video_generation

curl https://api.modelsite.ai/v2/video_generation \
  -H "Authorization: Bearer $MODELSITE_API_KEY" -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-H3",
    "content": [{ "type": "text", "text": "Epic space-opera trailer: a captain watches her fleet jump away through the observation window" }],
    "resolution": "768P",
    "ratio": "16:9",
    "duration": 5
  }'

Request body (OpenAI-multimodal-style content[], everything flat at the top level)

Field	Type	Notes
`model`	string (required)	`MiniMax-H3` (case-insensitive)
`content`	array (required)	Content items, below. A non-empty `text` item is required (the prompt)
`resolution`	string (required)	`768P` (standard) / `2K` (high definition)
`duration`	integer (required)	`[4, 15]` seconds
`ratio`	string	Aspect ratio; the rule depends on the job — below

content items

type	Fields	Notes
`text`	`text`	The prompt (required; the first text item wins)
`image_url`	`image_url.url` + `role`	Image: `role` is `first_frame` / `last_frame` (≤1 each; defaults to first frame when omitted) or `reference_image` (≤9). Frames and references are mutually exclusive
`video_url`	`video_url.url`	Reference video (≤3 clips, 2–15 s each, ≤15 s total, ≤50MB)
`audio_url`	`audio_url.url`	Reference audio (≤3 clips, 2–15 s each, ≤15MB)

Image limits: JPG/JPEG/PNG/WEBP/HEIC/HEIF, sides [256, 5760]px, aspect 0.4–2.5, ≤30MB.

ratio rules (per job)

Job	Rule
Text to video	Required, and must not be `adaptive`: `21:9` / `16:9` / `4:3` / `1:1` / `3:4` / `9:16`
Image to video	Always `adaptive` (follows the input image); other values are ignored
Multimodal reference	Optional, default `adaptive`; an explicit ratio is honored

Other vendor-native parameters (e.g. callback_url) go flat at the top level and pass through verbatim.

Create response

{ "task_id": "enc_v2_xxxxx" }

Polling

GET/v2/query/video_generation/{task_id}

curl "https://api.modelsite.ai/v2/query/video_generation/$TASK_ID" \
  -H "Authorization: Bearer $MODELSITE_API_KEY"

{
  "task": {
    "id": "enc_v2_xxxxx",
    "status": "succeeded",
    "content": { "url": "https://.../video.mp4" }
  }
}

Field	Notes
`task.id`	Job ID (the token returned at creation; pass it back verbatim)
`task.status`	Lowercase: `queued` / `running` / `succeeded` / `failed`
`task.content.url`	Video URL, only on success; download promptly
`task.error`	Only on failure: `{"code": "...", "message": "..."}`

This entrance has no delete endpoint — a job cannot be cancelled once created; wait for its terminal state.

Billing

Billed by actual generated seconds; 768P and 2K have different rates; failed jobs are not billed. Rates on the Models page.

Error handling

This entrance uses the MiniMax-native error shape, not the /api/v1 one:

{
  "type": "error",
  "error": { "type": "bad_request_error", "message": "...", "http_code": "400" }
}

Creation-time parameter errors return 400 + bad_request_error. Platform-level errors (auth, rate limits, balance) are covered in Errors.