> ## Documentation Index
> Fetch the complete documentation index at: https://docs.qingbo.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# MiniMax H3

> MiniMax H3 — omni-modal video generation. Text, image and reference inputs in one endpoint, native audio, 4–15 seconds.

MiniMax H3 is MiniMax's omni-modal video model, released July 2026. One endpoint takes **text, images, video and audio** as input and renders a finished clip **with its own audio track** in a single pass — no separate scoring step.

Three modes share the endpoint; what you send decides which one runs:

* **Text to video** — `prompt` only
* **Image to video** — `image_urls`, or `first_frame_image` / `last_frame_image` for an explicit first/last frame
* **Reference to video** — `video_urls` (camera work and style) or `audio_urls` (rhythm and score)

Async: the request returns a `task_id`; poll [task status](/en/api-reference/task/status) until `completed`.

## Pricing

<Note>
  Prices are not listed here — they follow upstream changes, and anything written into the docs drifts out of date.

  **Live prices:** the "Model Market" in the console, or the `GET /v1/models` endpoint (see [Models](/en/docs/models)).
  **What a call actually cost:** every response carries `cost` / `usage.cost` — that is the authoritative number.
</Note>

## Available models

| Model ID                  | Resolution      | Duration (sec)   | Audio    | Notes                                                                   |
| ------------------------- | --------------- | ---------------- | -------- | ----------------------------------------------------------------------- |
| `minimax-h3`              | `768p` / `2k`   | any integer 4–15 | ✅ native | Omni-modal input, finished clip with sound in one pass                  |
| `minimax-h3-max`          | `480p` / `768p` | any integer 4–15 | ✅ native | Faster, cheaper low-resolution tier with the same capabilities as H3    |
| `minimax-h3-regeneration` | `2k`            | any integer 4–15 | ✅ native | Re-renders an existing 768p H3 clip at native 2K; `video_urls` required |

Billed by **resolution × duration**; `2k` costs more. **A reference video (`video_urls`) is billed at 2× the unit price** — the upstream charges input-video seconds plus output seconds, and this is an interim approximation; `minimax-h3-regeneration`'s unit price already includes the input, so it is not doubled. The first 5 reference images are free; beyond that the upstream charges per image, which we currently **do not** pass on.

## Examples

<CodeGroup>
  ```bash Text to video theme={"system"}
  curl -X POST https://www.qingbo.dev/v1/tasks \
    -H "Authorization: Bearer $WAVE_API_KEY" \
    -H "Content-Type: application/json" \
    -d '{
      "model": "minimax-h3",
      "prompt": "a mountain town at dawn, mist lifting, camera pushing from the rooftops toward the ridge",
      "resolution": "2k",
      "duration": 8
    }'
  ```

  ```bash Image to video theme={"system"}
  curl -X POST https://www.qingbo.dev/v1/tasks \
    -H "Authorization: Bearer $WAVE_API_KEY" \
    -H "Content-Type: application/json" \
    -d '{
      "model": "minimax-h3",
      "prompt": "she slowly lifts her head, wind moves her hair, slight push in",
      "image_urls": ["https://cdn.example.com/portrait.jpg"],
      "resolution": "768p",
      "duration": 6
    }'
  ```

  ```bash First and last frame theme={"system"}
  curl -X POST https://www.qingbo.dev/v1/tasks \
    -H "Authorization: Bearer $WAVE_API_KEY" \
    -H "Content-Type: application/json" \
    -d '{
      "model": "minimax-h3",
      "prompt": "day turns to night, the lights come on one by one",
      "first_frame_image": "https://cdn.example.com/day.jpg",
      "last_frame_image": "https://cdn.example.com/night.jpg",
      "duration": 10
    }'
  ```

  ```bash Reference to video theme={"system"}
  curl -X POST https://www.qingbo.dev/v1/tasks \
    -H "Authorization: Bearer $WAVE_API_KEY" \
    -H "Content-Type: application/json" \
    -d '{
      "model": "minimax-h3",
      "prompt": "same camera rhythm, snowy mountains instead",
      "video_urls": ["https://cdn.example.com/reference.mp4"],
      "audio_urls": ["https://cdn.example.com/bgm.mp3"],
      "resolution": "2k",
      "duration": 12
    }'
  ```
</CodeGroup>

## Parameters

<ParamField body="model" type="string" required>
  Always `minimax-h3`
</ParamField>

<ParamField body="prompt" type="string" required>
  What the shot should show. Write it even when you supply references — the references say what it looks like, the prompt says what happens
</ParamField>

<ParamField body="resolution" type="string" default="768p">
  `768p` / `2k`
</ParamField>

<ParamField body="duration" type="integer" default="5">
  Any integer from 4 to 15 seconds
</ParamField>

<ParamField body="image_urls" type="string[]">
  Reference images, or the opening frame. Publicly reachable image URLs
</ParamField>

<ParamField body="first_frame_image" type="string">
  First-frame image URL
</ParamField>

<ParamField body="last_frame_image" type="string">
  Last-frame image URL; pair with `first_frame_image` for an A-to-B transition
</ParamField>

<ParamField body="video_urls" type="string[]">
  Reference video — supplies camera movement, pacing and style
</ParamField>

<ParamField body="audio_urls" type="string[]">
  Reference audio — shapes the rhythm and score of the result
</ParamField>

<ParamField body="callback_url" type="string">
  Called when the task reaches a terminal state; see [callbacks](/en/docs/sync-async#async-task-webhook)
</ParamField>

<Note>
  **Image mode and reference mode are mutually exclusive.** First/last frames (`first_frame_image` / `last_frame_image`) cannot be combined with reference media (`video_urls` / `audio_urls`) in one request — upstream treats them as two different generation modes.
</Note>

## Limits

| Item                     | Limit                  |
| ------------------------ | ---------------------- |
| Reference images         | up to 9                |
| Reference videos         | up to 3, each 2–15 sec |
| Reference audio          | up to 3, each 2–15 sec |
| Total reference duration | ≤ 15 sec               |
| Prompt                   | ≤ 7000 characters      |

## Related

* [Task system](/en/docs/task-system) — submit, poll, callbacks
* [Models](/en/docs/models) — list available models and live prices