How to call the MiniMax H3 video API (768P and 2K, native stereo audio)
To generate a MiniMax H3 video on RelayDance, POST to https://relaydance.com/v1/video/generations with model set to minimax-h3-768p or minimax-h3-2k, a prompt, a duration between 4 and 15 seconds, and an aspect ratio. Poll GET /v1/video/generations/{task_id} until the status is succeeded, then read the video url. Every clip ships with native stereo audio at no extra charge; 768P costs $0.10 per second (see relaydance.com/models).
Why MiniMax H3
MiniMax H3 is the strongest reason to care about audio-first video APIs right now: the soundtrack is generated natively together with the picture, not muxed on afterwards, so footsteps, ambience and dialogue stay in sync. The 768P tier renders 1344x768 at 24 fps, and the 2K tier targets maximum picture quality. Billing is strictly per second of generated output, so a 6-second clip costs exactly 6 seconds of the listed rate. It complements Seedance on RelayDance: same endpoint, same key, different strengths.
Set the base URL and authenticate
Point your OpenAI-compatible client at https://relaydance.com/v1 and pass your key as Authorization: Bearer YOUR_API_KEY. Create a key at the console (relaydance.com/console). No MiniMax account, enterprise contract or KYC is needed; payments are pay-as-you-go via card (Stripe) and USDT. The full request schema lives at relaydance.com/docs.
Text to video: your first request
The minimal request is a model, a prompt, a duration and a ratio. For text-to-video the ratio field is required (16:9, 9:16, 1:1); if you omit it, RelayDance defaults to 16:9. The value adaptive is rejected upstream unless a reference image is attached.
curl -X POST https://relaydance.com/v1/video/generations \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax-h3-768p",
"prompt": "a corgi surfing a small wave at sunset, cinematic, gentle waves, seagulls in the distance",
"duration": 6,
"ratio": "16:9"
}'
The response contains a task_id. Duration accepts integers from 4 to 15; the OpenAI-style string field seconds is also accepted.
Image to video and reference images
Attach a single reference image by adding "image": "https://example.com/photo.jpg" to the same request body. With a reference image attached, "ratio": "adaptive" becomes valid and follows the source image. For multiple reference images, pass a metadata.content[] array where each item is {"type": "image_url", "image_url": {"url": "..."}, "role": "reference_image"}. The first 5 reference images in a request are free; each additional image costs about $0.05.
Poll the task and read the result
Poll GET /v1/video/generations/{task_id} until status is succeeded or failed; the succeeded payload carries the video url. Generation typically completes within a few minutes (2K takes longer than 768P). Failed requests are never billed: the pre-authorized hold is released automatically.
Pricing
MiniMax H3 is billed per second of generated video. RelayDance places a worst-case hold when the task is submitted and automatically refunds the difference once the real output length is known. Live rates are published at relaydance.com/models.
| Option | Price | Example |
|---|---|---|
| minimax-h3-768p | $0.10 / second | 6-second clip = $0.60 |
| minimax-h3-2k | $0.1625 / second | 6-second clip = $0.975 |
| Extra reference images | about $0.05 / image after the first 5 | 7 images = about $0.10 extra |
Parameter reference
| Field | Meaning |
|---|---|
model | minimax-h3-768p or minimax-h3-2k; the output resolution is fixed by the model you pick |
prompt | Text description; describe sound cues too, the model generates the audio track natively |
duration | Integer seconds, 4 to 15 (string field seconds also accepted; default 5) |
ratio | 16:9, 9:16, 1:1; required for text-to-video, adaptive only with a reference image |
image | URL of a single reference image (alias image_url) |
metadata.content[] | Advanced: full content array with multiple reference_image items |
Steps to your first MiniMax H3 video
- Create an API key at https://relaydance.com/console.
- POST to
/v1/video/generationswith modelminimax-h3-768p, a prompt,duration4 to 15, andratio. - Optionally attach
imagefor image-to-video, ormetadata.content[]for multiple reference images. - Poll
GET /v1/video/generations/{task_id}until status is succeeded or failed. - Download the video from the returned url; the stereo audio track is already embedded.
FAQ
Is the audio track an extra charge? No. Native stereo audio is part of every MiniMax H3 generation on both tiers; you pay only the per-second video rate.
How long can a clip be? 4 to 15 seconds. For clips up to 30 seconds, use Seedance 2.5 (doubao-seedance-2-5-720p) on the same endpoint.
Am I charged when a request fails? No. Per the official documentation at relaydance.com/models, failed or errored requests are never billed.
768P or 2K? 768P renders 1344x768 at 24 fps and is the value pick at $0.10 per second. 2K is the quality pick at $0.1625 per second and takes longer to generate.
Do I need a MiniMax account? No. One RelayDance key covers MiniMax H3, Seedance and every other model in the catalog.
Per the official documentation at relaydance.com: "Failed or errored requests are never billed."
Per the official documentation at relaydance.com/docs: "Change the base_url to https://relaydance.com/v1 and keep using the OpenAI SDK."
Key facts and figures
| Item | Value | Source |
|---|---|---|
| MiniMax H3 768P price | $0.10 / second | relaydance.com/models |
| MiniMax H3 2K price | $0.1625 / second | relaydance.com/models |
| Audio | Native stereo audio on every clip, no extra charge | relaydance.com/models |
| Clip length | 4 to 15 seconds, integer | relaydance.com/docs |
| Reference images | First 5 free, then about $0.05 each | relaydance.com/models |
| API protocol | OpenAI compatible; base_url https://relaydance.com/v1 | relaydance.com/docs |
| Billing on failed requests | Failed or errored requests are never billed | relaydance.com/models |
Data last updated 2026-08-15; live prices are authoritative on the official /models page.