MiniMax Video API

This page specifies the MiniMax H3 video model minimax-h3. It shares the POST /v1/videos create and GET /v1/videos/{id} polling flow with the other video models; authentication, media upload, polling states, downloads and error handling are covered in the Complete Video API Standard. This page lists only the fields and rules specific to H3.

Updated 2026-10-13. H3 is sold in three tiers only: 768P, 1080P and 2K; 480P and 4K have been retired. New requests support polling only; do not send callback_url.

Text-to-video example

curl --request POST 'https://api.88api.ai/v1/videos' \
  --header 'Authorization: Bearer sk-xxxx' \
  --header 'Content-Type: application/json' \
  --data-raw '{
    "model": "minimax-h3",
    "prompt": "A silver sports car in a rainy city at night, low tracking shot, neon reflections, cinematic",
    "duration": 5,
    "resolution": "1080p",
    "ratio": "16:9"
  }'

Save the id from the response, poll GET /v1/videos/{id} until it completes, then download the video from the url in the completed response. minimax_h3 is another name for the same model, with identical usage and billing.

Parameters

H3 has a single model name. Choose the output tier with resolution (billing follows the tier) and the frame shape with ratio:

• 2K is produced by an upscaling workflow, not native generation.
• Do not pass the ratio in size. size only accepts a tier such as 1080p or dimensions such as 1920x1080; 16:9 is read as a resolution and rejected.
• Results always include an audio track; there is no generated-audio switch. The delivered duration is usually 0.1–0.6 seconds longer than requested.
• Invalid parameters (tier, ratio, duration, media count or combination) return 400 at submission; no task is created and nothing is billed.

Generation time and polling

H3 generation time depends on the request parameters:

• Regular requests (16:9 or 9:16, text-to-video, reference images, or 768P first and last frames) finish faster: about 2–3 minutes at 768P and about 5–8 minutes at 1080P or 2K.
• Other ratios, a first frame only, first and last frames at 1080P/2K and similar combinations take noticeably longer; with queuing they can need 10–60 minutes.

Set the polling timeout to at least 60 minutes; a polling interval of 10–30 seconds is enough.

Reference media

Put reference images in metadata.referenceImages and reference audio in metadata.referenceAudios. Up to 9 reference images and 3 reference audio clips; reference audio must be accompanied by at least one reference image.

{
  "model": "minimax-h3",
  "prompt": "The person from the reference image dances to the rhythm of the reference audio as the camera slowly circles",
  "duration": 6,
  "resolution": "768p",
  "ratio": "16:9",
  "metadata": {
    "referenceImages": ["https://cdn.example.com/dancer.jpg"],
    "referenceAudios": ["https://cdn.example.com/music.mp3"]
  }
}

First and last frames

Use metadata.firstFrame and metadata.lastFrame. They cannot be combined with reference images or reference audio. Both frames at 768P give the fastest results; you can also provide only a first frame, or use first and last frames at 1080P or 2K, but those requests take longer to generate.

{
  "model": "minimax-h3",
  "prompt": "The camera glides from the coast at dawn to the same bay at dusk",
  "duration": 5,
  "resolution": "768p",
  "ratio": "16:9",
  "metadata": {
    "firstFrame": "https://cdn.example.com/start.jpg",
    "lastFrame": "https://cdn.example.com/end.jpg"
  }
}

Top-level images

When top-level images (or input_reference) contains 2 images, they are read as the first and last frames. With 1 image or 3 or more, the purpose is not fixed (they may be treated as reference images or a first frame). Use metadata.referenceImages or metadata.firstFrame/lastFrame to state the purpose explicitly.

Media requirements

• Provide images and audio as publicly accessible HTTPS URLs or Base64 data URLs; you can also upload local image and audio files directly in a multipart form.
• Images are JPG, PNG or WebP, preferably no larger than 10 MB each; audio is MP3 or WAV, up to 20 MB per clip; inline Base64 media may total up to 32 MB.
• Do not use local file paths, private network addresses or browser blob: URLs.
• Result URLs expire (after at least 24 hours); download and save the video promptly once the task completes.