MiniMax H3 Video Gateway

Gateway Base URL: https://maasunion.com/minimax
Official reference: MiniMax Video Generation API

This page is for callers using the MaasUnion aggregation platform: how to submit a video generation task, poll the result, and troubleshoot errors.

Product Overview

MiniMax H3 is MiniMax's flagship long-form, multimodal video generation model. It supports:

• Text-to-video / Image-to-video: 4–15 seconds per generation
• Multimodal references: first/last frame, image reference, video reference, audio reference
• H3-Context-IR: expand short prompts into structured, semantically richer prompts (does not consume video quota)
• Video regeneration: upgrade 768P results to 2K, or regenerate from an existing 768P video

On MaasUnion, H3 is exposed through a dedicated gateway path and is independent from other video channels (Kling, Doubao, etc.)—task IDs, billing, and routing are fully isolated. Clients use a single Authorization: Bearer <TOKEN> header and do not need to know about the upstream.

Quick Start

1. Get a Token

Create a token in the MaasUnion console under Tokens. The token determines your identity, quota, and the models you can access.

2. Make Your First Call

Point any HTTP client at the MaasUnion gateway path, and use the token above for authentication:

curl --request POST "https://maasunion.com/minimax/v2/video_generation" \
  --header "Authorization: Bearer $NEW_API_TOKEN" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "MiniMax-H3",
    "content": [{"type": "text", "text": "Cinematic sunrise, camera slowly pushing in toward a mountain lake"}],
    "resolution": "768P",
    "duration": 6,
    "ratio": "16:9"
  }'

Response:

{"task_id": "424010985738629"}

3. Poll the Result

curl --request GET "https://maasunion.com/minimax/v2/query/video_generation/$TASK_ID" \
  --header "Authorization: Bearer $NEW_API_TOKEN"

On success, read the video URL from task.content.url. The URL is time-limited, so download or transfer the file to your own storage as soon as possible.

The official recommendation is to poll every 10 seconds. If you would rather not poll, you can pass callback_url when creating the task; the gateway will POST to that address whenever the status changes.

Endpoints

Every create endpoint returns a task_id; use the query endpoint to fetch the final result. Task records are retained for 7 days.

See the dedicated pages for parameter and field details: Create video · Context-IR · Video regeneration · Query tasks.

Task Status & Response Fields

Status progresses queued → running → succeeded / failed / cancelled.

Request Constraints

Content & duration

• model must be MiniMax-H3
• content must contain at least one non-empty text entry; max 7,000 characters
• Video duration 4–15 seconds; resolution 768P or 2K
• Supported ratios: adaptive, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16
• Text-to-video requires ratio and cannot be adaptive; image-to-video uses adaptive

Multimodal

• first_frame / last_frame cannot be combined with reference_image / reference_video / reference_audio
• Image ≤ 30 MB, video ≤ 50 MB, request body ≤ 64 MB
• For large files, use a public URL or mm_file://{file_id}

Tasks & expiry

• Task records are retained for 7 days
• Successful video URLs are time-limited; download or transfer immediately
• Recommended polling cadence: every 10 seconds, or use callback_url

Pricing

Charged in CNY and deducted from your MaasUnion quota. Failed or cancelled tasks are refunded per the standard rules.

See the official MiniMax pricing page for the latest rates. Context-IR does not produce video and only consumes token quota.

Troubleshooting

HTTP status codes

Business failure (status: failed)

When querying a failed task, task.error contains the code and message. Common cases:

Best Practices

• Polling: poll every 10 seconds, or use callback_url for lower latency
• Large files: prefer public URLs or mm_file://{file_id} to avoid base64 bloat
• Persist results: download or copy the video to your own storage as soon as the task succeeds
• Clean up: actively DELETE failed or unwanted tasks for easier auditing
• Multimodal isolation: do not mix first/last frame and reference assets; for character/scene consistency, put all assets in the same reference_image or reference_video group
• Prompt refinement: use Context-IR to expand your short prompt first, then copy content.prompt into the generation request for more stable results
• 2K upgrade: run 768P first to validate, then use video regeneration to upgrade to 2K to save cost