MiniMax H3

POST /v1/videos

The MiniMax H3 family is served through one endpoint; switching models only changes the model field. All five models share the same request shape, task polling, and response structure.

Model IDs

Model ID	Resolutions	Duration	Reference media	Free input images
`tk-minimax-h3`	768P, 2K	4–15 s	Images, videos, audio	First 5
`tk-minimax-h3-highspeed`	480P, 768P	4–15 s	Images, videos, audio	First 5
`tk-minimax-h3-lite`	480P, 768P	4–15 s	Images, videos, audio	First 5
`tk-minimax-h3-max`	480P, 768P	5–15 s	Images, videos, audio	First 2
`tk-minimax-h3-max-turbo`	480P, 768P	5–15 s	First/last frames only	—

Supported parameters

Checked against the API request validator on November 10, 2026.

API field	Type	Required	Values / limit
`model`	string	Yes	One of the model IDs above
`prompt`	string	Yes	Nonempty; at most 7,000 characters
`seconds`	integer	No	4–15 (`max`/`max-turbo`: 5–15); default `5`
`metadata.resolution`	string	Yes	`768P`, `2K` (h3) or `480P`, `768P` (other models)
`metadata.ratio`	string	See rules	`16:9`, `4:3`, `1:1`, `3:4`, `9:16`, `21:9`, `adaptive`
`metadata.media`	array	No	`{type, url}` entries; see reference media below
`metadata.extra`	object	No	`max`/`max-turbo` only; forwarded to the model (e.g. `prompt_expansion_mode`)

Aspect ratio rules

Text-to-video (no reference media): ratio is required and must not be adaptive.
First/last-frame generation: always adaptive; the output follows the input image. A client-supplied value is ignored.
Reference generation: defaults to adaptive; an explicit ratio is allowed. tk-minimax-h3-lite requires an explicit ratio (rejects adaptive).

Reference media

Each metadata.media entry contains type and url:

`type`	Limit	Notes
`first_frame`	1	Text and image content in the prompt is still required
`last_frame`	1	Requires `first_frame` in the same request
`reference_image`	9	Not supported by `tk-minimax-h3-max-turbo`
`reference_video`	3	Not supported by `tk-minimax-h3-max-turbo`; 2–15 s per clip, 15 s total
`reference_audio`	3	Not supported by `tk-minimax-h3-max-turbo`; 2–15 s per clip, 15 s total

Up to 12 reference files per request; first/last frames and reference media cannot be mixed.
JPEG/PNG/WebP images up to 30 MB; MP4/MOV videos; WAV/MP3 audio.
URLs must be directly downloadable (no login, cookies, or hotlink protection).
An image_url entry without a role counts as a reference image; on max-turbo such entries are rejected.

Request example

curl https://api.apimart.cc/v1/videos \
  -H "Authorization: Bearer $HIPIX_API_KEY" \
  -H "Content-Type: application/json" \
  --data-raw '{"model": "tk-minimax-h3", "prompt": "A cat running on the beach, cinematic light", "seconds": 5, "metadata": {"resolution": "768P", "ratio": "16:9"}}'

Results

Save the returned id and poll GET /v1/videos/{id}. On completed, url points to the HiPix content endpoint; you can also request /v1/videos/{id}/content. The response does not promise dimensions, duration, or a usage breakdown. Check usage logs for the actual charge.

Billing rules

Output video is billed per second, at the per-second rate of the requested resolution. Rates per model are shown on the model square.
Reference video input is billed too, by input duration: for tk-minimax-h3, -highspeed and -lite at the same per-second rate as the output resolution; for tk-minimax-h3-max at a separate (higher) input rate. Measured input seconds come from the completion response.
Input images: the first 5 input images per task are free on h3/highspeed/lite (first 2 on max); additional input images are billed per image. Input images include first/last frames and reference images; the billed count is the total input image count minus the free allowance.
Reference audio input is free. Generated audio is included in the per-second output price.
Failed tasks are not billed, including tasks rejected by content moderation.
Reservation is an estimate taken before submission (output duration plus the 15-second reference-video maximum when a reference video is present). After completion the charge is settled from the measured usage; a refund or additional charge applies the difference. When the completion response lacks measured reference-usage numbers, the task stays unresolved instead of settling from the estimate.

Model-specific notes

tk-minimax-h3 is the only model with 2K output; 480P/1K/3K on this model, and any 2K on the other models, are rejected.
tk-minimax-h3-max and tk-minimax-h3-max-turbo require at least 5 seconds.
tk-minimax-h3-max-turbo rejects every reference_* entry; images used as first_frame/last_frame must carry an explicit role.
metadata.extra (object, or a JSON string) is forwarded to the model unchanged and only accepted on max/max-turbo, e.g. {"prompt_expansion_mode": "quality"}.