# MiniMax H3 API

> Public documentation index for the MiniMax H3 asynchronous video-generation API.

## Start Here

- [API Reference](https://minimax.api.easyframe.cn/docs/): Human-readable request, response, media, and prompt documentation.
- [Model Catalog](https://minimax.api.easyframe.cn/v1/models): Public model list, availability, resolutions, pricing, and canvases.
- [Model Capabilities](https://minimax.api.easyframe.cn/v1/model-capabilities): Machine-readable request contract for every model.
- [Full LLM Reference](https://minimax.api.easyframe.cn/llms-full.txt): Complete public API contract in plain text.
- [API Studio](https://minimax.api.easyframe.cn/studio/): Browser client for testing the public API.

## Core Contract

- Base URL: `https://minimax.api.easyframe.cn`
- Authentication: `Authorization: Bearer <API_KEY>`
- Create a TOS material upload: `POST /v1/assets`
- Confirm a completed material upload: `POST /v1/assets/{asset_id}/complete`
- Create a video: `POST /v1/videos`
- Check a task: `GET /v1/videos/{id}`
- MiniMax official H3 V2 create compatibility: `POST /v2/video_generation`
- MiniMax official H3 V2 query compatibility: `GET /v2/query/video_generation/{task_id}`
- A completed status contains a one-hour presigned TOS `url` that downloads without an API key; `GET /v1/videos/{id}/content` remains a redirect-compatible fallback. Generated assets are retained for 24 hours after task completion.
- Cancel a task: `POST /v1/videos/{id}/cancel`
- One production queue episode has a 240-second supply deadline. An unclaimed
  task remains queued and forces capacity supply; waiting does not trigger a
  refund. `queue_timeout` is retained only for historical tasks.
- Ark image Base URL: `https://minimax.api.easyframe.cn/api/v3` (public paths never include the provider's internal `/plan` segment).
- Ark image generation: `POST /api/v3/images/generations` when the selected image model is enabled and explicitly priced.

## Important Constraints

- Public models: `minimax-h3-mini`, `minimax-h3-fast`, `minimax-h3-base`.
- Generation modes: `t2va`, `i2va`, `fl2va`, `l2va`, `ref2va`.
- Resolutions: `480p` or `720p`.
- Official V2 `MiniMax-H3` maps to Base and accepts `768P` as the public name
  for the production 720p canvas. `2K`, 4-second output, reference video,
  callbacks, and `mm_file://` are rejected before task creation or billing.
- Official image-to-video ratios are treated as adaptive. Unresolved adaptive
  requests use `16:9` and expose the adjustment in `X-H3-Adjusted`.
- Official V2 text items allow 7,000 characters each; their combined native
  prompt is limited to 30,000 characters.
- `seconds` is required and must be exactly one of: `5`, `6`, `7`, `8`, `9`, `10`, `11`, `12`, `13`, `14`, `15`.
- `prompt` is required and limited to 30000 characters.
- Send JSON for text-only requests and multipart form data when uploading media.
- JSON image/audio inputs may use a ready `asset_...` returned by the material
  upload API or a public HTTP(S) URL. URL, multipart, and base64 media are
  validated and staged into TOS by the platform. The execution service receives
  only fresh TOS download URLs, never the original external URL.
- Reference-video input is temporarily unsupported. A JSON or multipart `videos` field returns HTTP 422 with `reference_video_not_supported` before billing or task creation; generated MP4 output is unaffected.
- `doubao-seedream-5.0-lite` is adapter-ready but remains disabled until an administrator persists an explicit price.
- Ark image models are not selectable through Auto or the browser console.

Use the model-capabilities endpoint as the source of truth for allowed request fields, media counts, upload sizes, and model availability.