MiniMax H3 — video API
Use minimax/h3 for MiniMax H3 text-to-video, video from a first frame or from a first and last frame, and video from reference images, at 768p or 2K, through the Quotaflow /openai/v1/videos task API: create a task, store the returned id, and poll until the task completes. It uses the same API as the other video models; compare them on the Videos page.
MiniMax H3, Kling, Wan, Seedance, and Veo are different video model families. minimax/h3 is never used as a fallback for another family's model, and no other model stands in for it.
Model contract
| Property | Value |
|---|---|
| Model id | minimax/h3, as /models lists it |
| Endpoint | POST https://api.quotaflow.ai/openai/v1/videos |
| Status endpoint | GET https://api.quotaflow.ai/openai/v1/videos/{task_id} |
| Content endpoint | GET https://api.quotaflow.ai/openai/v1/videos/{task_id}/content |
durationSeconds | 4–15; default 5 |
resolution | 768p or 2k; default 768p |
aspect_ratio | text_to_video: 21:9, 16:9, 4:3, 1:1, 3:4, or 9:16; default 16:9. image_reference: the same, or adaptive (the default). Frame modes: ignored, the frame decides |
| Modes | text_to_video (the default), image_to_video, first_last_frame, image_reference |
first_frame_url, last_frame_url | Public https:// image links |
reference_image_urls | 1–9 public https:// image links |
max_cost_usd | Required |
Call /models with your production key before launch. Video access is key-enabled: the default all-model key has it. Only a key with access to the MiniMax H3 video line lists minimax/h3.
Prices
Prices are USD per second of video, by the output resolution. Every mode is charged at the same rate. The seconds charged are the input video seconds plus the output video seconds.
| Unit | 2k | 768p |
|---|---|---|
| Video, per second (input plus output) | $0.13 | $0.08 |
| Input image beyond the first 5 per task, each | $0.04 | $0.04 |
The first 5 input images of a task (frames and reference images together) are free; each further one is $0.04 at either resolution. The max_cost_usd hold taken when the task is created is separate from the final charge.
Text-to-video example
curl https://api.quotaflow.ai/openai/v1/videos \
-H "Authorization: Bearer $QUOTAFLOW_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax/h3",
"prompt": "A paper boat drifting across a still pond at sunrise, slow camera pan.",
"durationSeconds": 5,
"resolution": "768p",
"aspect_ratio": "16:9",
"max_cost_usd": 2
}'
First-frame example
Set mode to image_to_video and send first_frame_url. For first_last_frame, also send last_frame_url. Frames are public https:// links; a Base64 data URL is refused. The video takes its shape from the frame, so aspect_ratio is ignored with a frame.
curl https://api.quotaflow.ai/openai/v1/videos \
-H "Authorization: Bearer $QUOTAFLOW_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax/h3",
"mode": "image_to_video",
"prompt": "Keep the product shape and color consistent, slow camera push-in.",
"first_frame_url": "https://example.com/frames/product.png",
"durationSeconds": 5,
"resolution": "768p",
"max_cost_usd": 2
}'
Reference-image example
Set mode to image_reference and send up to 9 public https:// image links in reference_image_urls, with no frames.
curl https://api.quotaflow.ai/openai/v1/videos \
-H "Authorization: Bearer $QUOTAFLOW_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax/h3",
"mode": "image_reference",
"prompt": "The character from the references walks through a rainy street at night.",
"reference_image_urls": ["https://example.com/refs/character-front.png", "https://example.com/refs/character-side.png"],
"durationSeconds": 5,
"resolution": "768p",
"max_cost_usd": 2
}'
Poll the task
curl https://api.quotaflow.ai/openai/v1/videos/$TASK_ID \
-H "Authorization: Bearer $QUOTAFLOW_API_KEY"
Download the completed video
curl -L https://api.quotaflow.ai/openai/v1/videos/$TASK_ID/content \
-H "Authorization: Bearer $QUOTAFLOW_API_KEY" \
--output h3.mp4
The content endpoint redirects to the video file; add ?format=json to receive { "url", "expires_at" } instead.
Notes
- The mode and the inputs must agree:
text_to_videotakes no frames and no references,image_to_videoexactly a first frame,first_last_frameboth frames, andimage_referenceonlyreference_image_urls. A mismatch is refused with400, as is amodeH3 does not have, such asvideo_edit.video_referenceis not served onminimax/h3: no capacity has passed its check. reference_image_urlstakeshttps://links only; anasset://id is refused with400.watermarkis passed to the model.callback_urlis accepted; anhttp://one is refused with400.generate_audio: false,return_last_frame: true,execution_expires_after, aframe_rateother than 24,quality,reference_shape, andreference_digestare refused with400.durationSeconds: -1is not accepted: send a length within 4–15 seconds, or omit it for 5.- If the selected key does not list
minimax/h3from/models, do not retry as a transient error; request MiniMax H3 video access for that key.
Next steps
- Read the Videos endpoint guide.
- Check
/modelswith your production key for currently enabled video models. - Browse all supported models.