MiniMax H3 Video Generation
The MiniMax H3 series uses the MiniMax native async API (/api/v2), alongside the Volcengine Seedance shape (/api/v3). Both follow "create task → poll status → read result", but differ in paths, field names, and billing.
MiniMax video is billed per second (per-second price depends on resolution), not per token.
Models and resolutions
Each H3 variant supports different resolutions. The available tiers are determined by the pricing configured for the model:
| Model | Supported resolutions | Notes |
|---|---|---|
| MiniMax-H3 | 768P / 2K | Standard, supports 2K |
| H3-Highspeed | 480P / 768P | Speed first |
| H3-Lite | 480P / 768P | Cost first |
| H3-Max | 480P / 768P | Quality first, minimum 5 seconds |
Passing an unsupported resolution returns 400 immediately without submitting to the upstream.
Key parameters
| Parameter | Type | Required | Description |
|---|---|---|---|
model | string | Yes | Model call name from the Model Plaza or model list |
content | array | Yes | Input content; must contain at least one non-empty text item |
resolution | string | Yes | 768P / 2K or 480P / 768P depending on the model |
duration | integer | Yes | Output duration in seconds, 4–15 (H3-Max minimum 5) |
ratio | string | Conditional | adaptive / 21:9 / 16:9 / 4:3 / 1:1 / 3:4 / 9:16; required for text-to-video and cannot be adaptive |
callback_url | string | No | Terminal-state callback URL (platform POSTs the result when the task finishes) |
aigc_watermark | boolean | No | Whether to add a watermark, default false |
content items
| Field | Type | Description |
|---|---|---|
type | string | text / image_url / video_url / audio_url |
text | string | Required when type=text |
image_url.url | string | Public URL |
role | string | first_frame / last_frame / reference_image / reference_video / reference_audio |
First/last frame (
first_frame+last_frame) and reference generation (reference_*) cannot be mixed.
Endpoints
| Operation | Method | Path |
|---|---|---|
| Create task | POST | /api/v2/video_generation |
| Query task | GET | /api/v2/query/video_generation/{task_id} |
| List tasks | GET | /api/v2/query/video_generation |
| Cancel/delete | DELETE | /api/v2/video_generation/{task_id} |
Task status: queued → running → succeeded / failed / cancelled.
Text to video
curl -X POST https://www.silvamux.com/api/v2/video_generation \
-H "Authorization: Bearer $SILVAMUX_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "<MODEL_CALL_NAME>",
"content": [{"type": "text", "text": "A golden retriever running on the beach, cinematic"}],
"resolution": "768P",
"duration": 5,
"ratio": "16:9"
}'
Successful creation responds:
{"task_id": "462160194171102"}
Query a task
curl "https://www.silvamux.com/api/v2/query/video_generation/<TASK_ID>" \
-H "Authorization: Bearer $SILVAMUX_API_KEY"
A successful response wraps the payload in task:
{
"task": {
"id": "462160194171102",
"model": "<MODEL_CALL_NAME>",
"status": "succeeded",
"resolution": "768P",
"duration": 5,
"content": { "url": "https://..." },
"usage": { "total_seconds": 5, "output_seconds": 5 }
}
}
- The artifact URL lives at
task.content.url, expires, so download it promptly. - Billed seconds are in
task.usage.total_seconds.
First-frame and reference generation
{
"model": "<MODEL_CALL_NAME>",
"content": [
{"type": "text", "text": "The camera pulls back slowly as steam rises"},
{
"type": "image_url",
"image_url": {"url": "https://example.com/first.png"},
"role": "first_frame"
}
],
"resolution": "2K",
"duration": 5,
"ratio": "adaptive"
}
Reference generation (reference_image / reference_video / reference_audio) supports up to 9 images, 3 videos, and 3 audio clips; audio cannot appear alone.
Cancel a task
curl -X DELETE "https://www.silvamux.com/api/v2/video_generation/<TASK_ID>" \
-H "Authorization: Bearer $SILVAMUX_API_KEY"
Cancelling a queued task is free (the pre-authorised amount is fully refunded). The upstream does not allow cancellation once a task is running.
Result callback
When callback_url is provided, the platform POSTs the result to that address on terminal state; the body matches the query API ({"task": {...}}).
Your callback URL is proxied by the platform, so you do not need to implement MiniMax's official challenge handshake.
Without callback_url you can poll the query API yourself; the platform also runs an internal polling fallback.
Billing
Billed by output video seconds, with a per-second price that varies by resolution tier. For example:
- 768P 5 seconds × unit price
- 2K 5 seconds × unit price (higher tier)
See the pricing table on the model detail page or the console pricing page for actual unit prices.
Differences from Volcengine Seedance
| Item | MiniMax H3 | Volcengine Seedance |
|---|---|---|
| Create path | /api/v2/video_generation | /api/v3/contents/generations/tasks |
| Create response | {"task_id": ...} | {"id": ...} |
| Query response | {"task": {...}} | Bare task object |
| Artifact field | task.content.url | content.video_url |
| Billing | Per second (by resolution) | Per token |
The two APIs coexist and their tasks are isolated: the /api/v3 task list excludes MiniMax tasks and vice versa.