MiniMax H3 API: Text to Video with Native Audio
Call the open-weight MiniMax H3 model over a hosted API: a prompt or a start frame in, a 4 to 15 second MP4 with synchronised stereo audio out, billed per second of video. Run the preset directly on this page or call the same agent-ready endpoint from your product.
Useful defaults, typed options.
The tool slug stays stable while your agent supplies named media inputs and a narrow set of documented options.
- No GPU to rent, no weights to download
- Flat per-second price, reserved before the render starts
- Same API and MCP server as the rest of the catalog
What this preset does
- 01Send a prompt, or a start frame to animate, to the hosted MiniMax H3 endpoint
- 02Render picture and stereo audio in one pass on a GPU worker
- 03Tag the output as AI-generated (EU AI Act art. 50)
- 04Return a playable MP4 with a download URL
Stable endpoint
POST /api/v1/tools/minimax-h3Know the boundaries before you run.
MiniMax H3 renders 4 to 15 seconds per call, at 480p or 768p, 24 fps, H.264 with AAC stereo
For image-to-video, pass image_url as the first frame and optionally last_image_url; the clip then follows the image aspect ratio
A render takes several minutes, longer at 768p than at 480p, and longer again when no worker is already warm. Poll the job or register a webhook
Priced per second of video: 5 credits at 480p, 6 at 768p, reserved when the job is queued and refunded if it fails
Prompts and outputs are moderated, and the MiniMax H3 Community License applies to what you generate
About minimax h3 api
An open-weight text-to-video model that generates picture and sound together, rather than adding audio afterwards. It renders 4 to 15 second clips at 480p or 768p, 24 fps, and also accepts a start frame for image-to-video.
6 credits per second of video at 768p and 5 at 480p, so a 6 second clip is 36 or 30 credits. The price is reserved when the job is queued, never changes with render time, and is refunded in full if the render fails.
No. The weights are open, so you can run them yourself on a rented GPU, but this endpoint hosts the model for you: an HTTP call or an MCP tool call, and an MP4 comes back. You pay per second of video rather than per hour of GPU.
Yes, subject to the MiniMax H3 Community License and our terms: no real-person likeness without consent, nothing unlawful, and you keep the AI-generated marking on published synthetic media. Every file we return carries that tag in its container metadata.
Yes. Pass image_url for the first frame, and last_image_url as well if you want the clip to land on a specific end frame. Starting from a frame you control gives more predictable motion than a prompt alone.
Connect the MCP server and call generate_video with a prompt, seconds, resolution and aspect_ratio, then get_job until it succeeds. Claude, Claude Code, Cursor, Codex and n8n all work; the connect page has the one-line command for each.