Skip to main content
MiniMax H3 is MiniMax’s omni-modal video model: one model takes text, images, video, and audio as input and returns video with its own music and sound effects. LaoZhang API bills it by video length, $0.05 per second at 768P, and any standard token group can call it without a dedicated video group.
MiniMax H3 is available only through the native /hailuo/v2 endpoints above. Requests sent to /v1/videos or /v1/video/generations are rejected and not charged.
MiniMax H3 text-to-video sample: an orange cat walking across a wet rooftop at dusk with city lights behind it

Output of the text-to-video example below: 768P, 5 seconds, 16:9

Create a token

In Token management, add a token with these settings: You don’t need a MiniMax account or a dedicated video group.

Pricing

MiniMax H3 is billed by video length. Prices from October 24, 2026: Official prices are from MiniMax pay-as-you-go pricing (MiniMax-H3, checked October 24, 2026). Landscape, portrait, and square videos cost the same; only resolution and seconds matter. How the billed seconds are counted:
  • Output: the duration you request. Clips usually run 0.1–0.5 seconds longer than requested at no extra cost.
  • Reference video: its length, rounded up to the next whole second; a 5.8-second clip counts as 6 seconds.
  • Reference images: images beyond the first 5 in one task are charged individually, and first-frame and last-frame images count toward the total.
Example costs for a single video:
  • 768P, 5-second text-to-video: 5 × $0.05 = $0.25
  • 2K, 10-second text-to-video: 10 × $0.08 = $0.80
  • 768P, 4 seconds with a 6-second reference video: (4 + 6) × $0.05 = $0.50
When you create a task, duration × rate is reserved from your balance, together with any images beyond the first five. For tasks with a reference video, the reference seconds are charged once the task finishes, so the call logs show a second settlement entry. Failed tasks are fully refunded, and requests rejected at submission are not charged. Resubmitting creates a new task that is billed again, so keep a mapping from your business ID to the task_id. The call logs show the exact charge for every task.

How it works

1

Create a task

Send POST /hailuo/v2/video_generation with a JSON body containing:
  • model: MiniMax-H3
  • content: one text item plus optional image, video, or audio inputs
  • resolution, duration, and ratio
Save the task_id from the response.
2

Poll the task

Call GET /hailuo/v2/query/video_generation/{task_id} every 10 seconds and read task.status. A 4–5 second 768P video usually finishes in about a minute; 2K and reference-video tasks take longer. Set an overall client timeout of 30 minutes.
3

Download the video

When task.status is succeeded, download the MP4 from task.content.url. The link stays available for 7 days and doesn’t need an Authorization header.

Set up

The cURL examples use jq to read fields from responses (install it with brew install jq on macOS). Set your key and base URL first:

Generate a video from text

A successful request returns only the task ID:
Text-only requests need a fixed ratio; adaptive isn’t accepted without an image.

Animate from a first frame or first and last frames

Use first_frame to set the opening shot, and add last_frame to set the closing shot as well. With ratio set to adaptive, the output follows the first frame’s aspect ratio:
Images can be public HTTPS URLs or Base64 data URLs such as data:image/jpeg;base64,.... For a first frame only, drop the last_frame item.

Generate from reference images, video, and audio

References let you put a specific character or object in the scene, reuse the motion and camera work of a clip, or follow the rhythm of an audio track. Refer to them in the prompt by type and order: the first reference image is <Picture 1>, the first reference video is <Video 1>, and the first reference audio is <Audio 1>.
When combining references:
  • Mix reference images, videos, and audio freely, but audio can’t be the only reference; add at least one reference image or video.
  • First and last frames can’t be used in the same request as references.
  • Video and audio inputs must be public HTTPS links the service can download directly.
  • Reference video is billed per second; see Pricing.

Poll and download

Get the task

Run it every 10 seconds until task.status is succeeded or failed. A successful response:
usage holds the billed seconds for the task: input_seconds is reference video, output_seconds is the generated video.

Download the video

The link stays available for 7 days, so copy the file to your own storage soon after the task finishes.

Minimal Python client

The script uses only the Python standard library. Set LAOZHANG_API_KEY before running it. It polls every 10 seconds with a 30-minute overall timeout and saves the result as h3-output.mp4.

API reference

Request basics

Request body

adaptive follows the image’s aspect ratio and can’t be used for text-only or audio-only requests. At 16:9, 768P output is 1344×768 and 2K output is 2560×1440.

content[] inputs

With a single image you can omit role and it’s treated as the first frame. With two or more images, set role on each.

Task status

At busy times a task can stay queued for a while and starts automatically in submission order. A task that hasn’t started after 30 minutes moves to failed and is refunded.

Errors

Some validation messages are returned in Chinese, as shown above. In production, store the task_id from the create response and the final task details so you can reconcile charges or contact support (hi@laozhang.ai).

FAQ

How much does a 5-second video cost?

$0.25 at 768P and $0.40 at 2K. Cost scales linearly with length, so 15 seconds at 768P is $0.75. Reference video adds its seconds at the same rate, and each reference image beyond the first five adds $0.03; see Pricing.

Should I use 768P or 2K?

Iterate on the prompt and composition at 768P with 4–5 second clips, then render the final version at 2K or a longer duration. 2K costs $0.03 more per second and takes longer to generate, so it’s best kept for footage meant for large screens or further editing.

Can I call MiniMax H3 through the OpenAI-style /v1/videos endpoint?

No. First and last frames, multiple reference images, reference video, and reference audio all depend on the content[] structure, so LaoZhang API exposes MiniMax H3 only on the native /hailuo/v2 endpoints. Requests to /v1/videos or /v1/video/generations are rejected without charge.

Do the videos have sound? Can I turn it off?

Every video comes with a stereo soundtrack whose music and effects follow the prompt and any reference audio. There’s no parameter to disable it; strip the audio track after download if you don’t need it.

Can I resubmit after a timeout or failure?

Check the task first. A client-side timeout doesn’t mean the task failed, so keep polling the same task_id; resubmitting creates a new task and bills it again. Only recreate the request after task.status is failed, and a failed task has already been refunded. 7 days. After that the link stops working and the task can’t be downloaded again, so save the file when the task finishes.