Skip to main content

​
Overview

MiniMax H3 (Hailuo 3.0) is an omni-modal video model released by MiniMax on 2026-07-31 (UTC+8). A single model takes in text, images, video, and audio and outputs video with a stereo audio track. APIYI serves MiniMax-H3 from a self-hosted deployment of the open weights, at 768P, 4–15 seconds per clip, billed per second.
Highlights: one endpoint covers text-to-video, first-frame / last-frame / first-and-last-frame video, and mixed image / video / audio reference generation. Limits are 9 reference images, 3 reference videos, and 3 reference audio clips, with at most 12 media items overall. Output includes a stereo audio track. Output video and input reference video each cost $0.03 per second (the official MiniMax API charges $0.08 for each). A 10-second text-to-video or image-to-video task without a reference video costs $0.30. Failed tasks are refunded automatically.

Video Generation API Reference

Create a task and query it by task_id, with Python / cURL / Node.js examples and a live Playground

Top-up Bonuses

Top-up bonuses lower the effective price further

​
Let an AI Agent Integrate It for You

Copy this prompt and paste it into Codex / Claude Code / Cursor

The agent first fetches the plain-text version of this page (append .md to any docs URL), then writes code for your stack. The common mistakes are already spelled out: the path needs /hailuo, results sit inside task, and duration must be an integer from 4 to 15.

Integrate or debug MiniMax-H3 video generation (text-to-video / first-and-last-frame video / reference video) in this project.

Read the docs before writing code: fetch https://docs.apiyi.com/en/api-capabilities/minimax-h3/overview.md for the plain-text version of this page; parameters and code samples are at https://docs.ap …

  1. Endpoints: submit with POST https://api.apiyi.com/hailuo/v2/video_generation, query with …

10 requirements in total. Use the button below to copy the full prompt.

​
Why APIYI’s MiniMax-H3?

Multiple generation modes

Text, first/last-frame, and mixed image / video / audio reference generation are available, with at most 12 media items overall; see Technical Specifications for the limits per media type

Per-second billing, refunds on failure

Output video and input reference video each cost $0.03 per second; failed tasks are refunded in full and submission errors are free

Top-up bonuses stack

Combine with top-up bonuses for a lower effective cost

Global access, no barriers

Connect directly to api.apiyi.com with one API Key; no overseas account needed

Full video model lineup

The same key also works with Seedance 2.0 / 2.5, Wan3.0, Wan2.7, and more

Professional support

Contact support with integration questions; enterprise customers get hands-on onboarding

​
Key Features

Native stereo audio

Output includes a stereo audio track driven by the prompt and reference audio, with support for generated sound and reference-audio reuse

7 aspect ratios

Six fixed ratios from 21:9 to 9:16, plus adaptive to follow the reference image

Any whole length from 4 to 15 s

Choose the output duration in whole seconds; reference video is billed separately by its input duration

Long prompts

Up to 7000 characters per prompt, enough for shot-by-shot descriptions

First/last-frame control

Provide only the first frame, only the last frame, or both to set how the clip starts and ends

Multiple reference images

Up to 9 reference images; tag characters and objects in the prompt with <Picture 1> and so on

Motion transfer from video

Up to 3 reference videos to copy camera moves and motion rhythm

Audio-driven video

Up to 3 reference audio clips; the picture follows the music or voice

​
Pricing

This channel runs MiniMax’s open H3 weights on our own deployment. It is not a relay of the official MiniMax API, so it has its own pricing:
This channel is a self-hosted deployment of the open weights, priced independently of the official MiniMax API, and prices may change; the table above is for reference only, and the Model Pricing tab in the top navigation is authoritative: Model Pricing. Official prices from platform.minimax.io/docs/guides/pricing-paygo (retrieved 2026-09-29).
Billing details:
  • Output video: $0.03 per second; input reference video: $0.03 per second. Final billing uses the upstream-reported output and input video usage; input video duration comes from usage.input_seconds
  • On acceptance, the platform pre-charges the output fee using the requested duration. After completion, it settles actual usage and charges or refunds the difference
  • A settlement log records only the adjustment, not the full cost again. Final cost = pre-charge + additional charges − refunds; a task does not always have exactly two billing records
  • Reference images (up to 9) and reference audio have no extra charge
  • Failed tasks (media download failure, unsupported format, execution failure, etc.) are refunded in full automatically
  • Requests that return 4xx / 5xx at submission are not billed; querying and downloading are free
  • See top-up bonuses for a lower effective cost
Examples at the standard rates above: These examples illustrate what is billed. Prices may change; check Model Pricing in the top navigation for the latest rates. Use the upstream-reported usage.input_seconds for the input duration, rather than the requested output duration.

​
Group Setup

MiniMax-H3 works in the default group, and the svip group works too; no dedicated group is needed. We recommend setting the Token’s billing model to Pay-as-you-go Priority. If a call returns “no available channels for the current group”, the Token’s group does not include this model or the model value is misspelled (it is case-sensitive).

​
Technical Specs

The official MiniMax H3 model supports 2K, but this channel only offers 768P; 2K is rejected.

​
Estimated Generation Time

Generation time for MiniMax-H3 depends mainly on two things: output duration and whether a reference video is included. The table below reflects the current state of this channel, based on 400+ successful tasks from 2026-09-29 to 2026-10-10 (UTC+8):
  • Time grows faster than duration: a 15-second clip takes about 5× as long as a 5-second clip, not 3×
  • Reference videos take the longest: the model parses the reference video frame by frame and computes it together with the new frames, multiplying the workload
  • Set client timeouts from this table: 30 minutes for tasks without a reference video and 2 hours for tasks with one, so running tasks are not mistaken for failures
These are this channel’s current measured numbers. For time-sensitive workloads, start with 4–6 second clips or use reference images instead of reference videos. A faster, higher-quality MiniMax group is coming soon; we will announce it on this page and in the Changelog.

​
API Endpoints

Primary host https://api.apiyi.com, backup host https://b.apiyi.com, same paths. Note that the paths start with /hailuo, not /v1.

​
Generation Modes

The generation mode is inferred from the media in content[]:

​
Referring to media in the prompt

Each media type is numbered by its order in content[]: the first and second reference images are <Picture 1> and <Picture 2>, the first reference video is <Video 1>, the first reference audio clip is <Audio 1>. For example:
  • First/last frames cannot be combined with any reference media
  • With a single image you can omit role (it is treated as the first frame); with two or more images, set role on each
  • The combined length of reference videos cannot exceed 15 seconds, otherwise the task fails (and is refunded); a single clip over 15 seconds is trimmed to its first 15 seconds

​
Reference-audio reuse and limits

Send audio with type: "audio_url", audio_url.url, and role: "reference_audio". It can be combined with images using role: "reference_image". Use <Audio 1>: fully_copy in the text field of a type: "text" item in content[] to guide audio reuse. The following is an excerpt from the prompt’s retention_analysis section; send it with the reference audio and a complete visual description:
fully_copy is an audio-reuse instruction in the prompt, not a separate JSON parameter or an original-audio passthrough switch. Audio reuse has been tested on this channel, but the output is generated by the model. Sample-for-sample identity, no added sound, and strict silence after the reference ends are not guaranteed. If the original soundtrack must remain entirely unchanged, replace the audio track after video generation and match the durations yourself.

​
Best Practices

1

Test with 4–5 seconds first

Billing is per second, so confirm composition and style with a short clip before rendering 10–15 seconds
2

Use a fixed ratio for text-to-video

16:9 for landscape, 9:16 for portrait, 21:9 for widescreen; with a first frame, adaptive keeps the image’s ratio
3

Host media on stable public storage

Use direct links from your own object storage or CDN so hotlink protection or expired signatures don’t break media downloads
4

Describe camera moves and sound

Cover the subject, action, camera movement, lighting, and the music and sound effects you want; the model generates them together
5

Poll every 10 seconds

A 5-second clip usually takes about 3 minutes; set the overall client timeout to 30 minutes without a reference video and 2 hours with one (see “Estimated Generation Time”)
6

Handle idempotency yourself

Keep a mapping from business ID to task_id; if a submission times out, look for an existing task before submitting again
7

Store the video right away

Download task.content.url with GET and serve it from your own storage

​
Error Codes & Retries

Client tips: set the submit timeout to 60 seconds (at peak times submission alone can take more than 10 seconds); only retry HTTP 500s and network errors, with backoff. Every run-stage failure is refunded automatically, and resubmitting creates a new charge.

​
FAQ

The path is missing the /hailuo prefix. The correct paths are /hailuo/v2/video_generation and /hailuo/v2/query/video_generation/{task_id}.
This channel supports 768P only. The official MiniMax model supports 2K, but this channel does not offer it.
No. duration is an integer from 4 to 15.
Output includes a stereo audio track, and there is no parameter to disable audio generation. Passing generate_audio=false returns HTTP 400 unsupported field generate_audio. Strip the track in post-processing if you don’t need it.
This endpoint does not support the seed parameter; passing it returns HTTP 400 unsupported field seed. Task queries do not return an actual seed for reproducibility either.
The current create and query endpoints do not return a model revision or a complete execution-parameter snapshot. Queries may include metadata such as model, duration, ratio, and resolution, but MiniMax-H3 is only a model name and does not identify the deployed weight snapshot.
Not at the moment. Resubmitting with the same Idempotency-Key still creates a new task that is billed separately. Track submitted tasks in your own application.
No. Once a task reaches failed it is refunded in full automatically; requests rejected at submission are not billed either.
No. All media must be public HTTPS URLs. Upload to your own object storage first and pass the link.
It follows the input image’s aspect ratio, e.g. 768×768 for a square image and 1344×768 for a 16:9 image. Text-only and audio-only requests cannot use adaptive.
A single clip over 15 seconds is trimmed to its first 15 seconds automatically, but if several reference videos add up to more than 15 seconds the task fails (and is refunded).
About 3 minutes for a 5-second clip and about 13 minutes for a 15-second clip; tasks with a reference video take much longer, about 85 minutes at 15 seconds. See “Estimated Generation Time” above for the full table.
We have not seen it expire quickly, but long-term availability is not guaranteed, so download and store the video promptly. Check the URL with GET; HEAD requests return 403.
First confirm the request body follows the rules (exactly one text item, no extra fields, https media links). If it does, it is usually transient overload: wait a few seconds and retry. Failed submissions are not billed.