# MiniMax H3

MiniMax H3 generates 2K video from four kinds of input: a text prompt, a starting image, a pair of first and last frames, or reference material. Output is mp4, five to 15 seconds, in aspect ratios from 21:9 through 9:16.

- **Model ID:** `minimax/minimax-h3`
- **Type:** video
- **Providers:** minimax
- **Context window:** Not applicable
- **Maximum output tokens:** 0
- **Pricing:** See the provider details below
- **Canonical page:** https://vercel.com/ai-gateway/models/minimax-h3

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { experimental_generateVideo as generateVideo } from 'ai';

const result = await generateVideo({
  model: 'minimax/minimax-h3',
  prompt: 'A serene mountain lake at sunrise.'
});
```

## About

MiniMax H3, released July 30, 2026, is MiniMax's video generation model. It takes a text prompt, a starting image, a pair of first and last frames, or reference material, and returns mp4 at 2K resolution.

Four generation modes cover most production needs. Text-to-video builds a clip from a description alone. First-frame image-to-video animates a still and follows that image's aspect ratio. First-to-last keyframe transitions define both endpoints and let MiniMax H3 generate the motion between them. Multimodal reference-to-video conditions a generation on reference images, video, or audio in a single request. Reference and keyframe modes are mutually exclusive, so choose one per request.

Reference mode uses ordering to hold a subject steady. Pass source material through `inputReferences` and refer to each item by its position in the prompt, such as `Image 1`. That gives you a way to carry a character or product across several clips.

Clips run from five to 15 seconds at 2K resolution in mp4. Aspect ratios include 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16, or the output adapts to a supplied image. Pick the ratio at request time instead of cropping afterward.

Call MiniMax H3 with `generateVideo` from the AI SDK, imported as `experimental_generateVideo`. Set the model to `minimax/minimax-h3`, pass your prompt plus optional `aspectRatio` and `duration`, and add `inputReferences` when you condition on source material. Generation runs longer than a text completion, so raise `pollTimeoutMs` for longer clips.

AI Gateway routes MiniMax H3 across MiniMax, tracks usage and cost per request, and handles retries and failover. The listed video rate is N/A, and the pricing panel on this page shows the current rate and unit. AI Gateway reflects provider pricing with no markup and charges no platform fee on inference.

## What to consider

Reference mode and keyframe mode can't be combined in one request, so decide which control you need before you build the call. Clip length caps at 15 seconds, so longer sequences take several generations stitched together downstream. Generation runs long enough that default timeouts cut it short, so raise `pollTimeoutMs` and extend fetch timeouts in Node.js. When you pass a starting image, the output follows that image's aspect ratio, so crop the source to the ratio you want to deliver.

## When to use

### Best For

- Text-to-video generation where the whole clip comes from a written description
- Product and character animation that starts from an existing still image
- Defined transitions where you supply the first and last frames and want the motion between them
- Multi-clip sequences that hold a subject steady through reference images, video, or audio
- Delivery pipelines that need 2K mp4 in vertical, square, and widescreen ratios from one model

### Consider Alternatives When

- A single clip has to run longer than 15 seconds
- You want reference conditioning and keyframe endpoints in the same request
- Delivery requires resolution above 2K
- The workload is text or code rather than video. MiniMax M3 covers reasoning and engineering tasks

## Best for

- **Text-to-Video Generation:** Clips built entirely from a written scene description
- **Still Image Animation:** Product or character stills animated from a single starting frame
- **Defined Frame Transitions:** Motion generated between first and last frames you supply
- **Consistent Multi-Clip Subjects:** Reference images, video, or audio that carry a subject across shots
- **Multi-Ratio Delivery:** Vertical, square, and widescreen 2K output from one model

## Consider alternatives

- **Clips Beyond 15 Seconds:** Longer sequences take several generations stitched together downstream
- **Reference Plus Keyframes:** The two modes are mutually exclusive, so split them across separate requests
- **Resolution Above 2K:** Output caps at 2K, so plan an upscaling step for larger delivery formats
- **Text and Code Workloads:** MiniMax M3 covers reasoning and engineering tasks instead of video

## Frequently asked questions

### What can MiniMax H3 generate video from?

A text prompt, a starting image, a pair of first and last frames, or reference material. Reference mode accepts images, video, or audio in a single request.

### What resolution and duration does MiniMax H3 output?

Output is mp4 at 2K resolution, from five to 15 seconds. Supported aspect ratios include 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16, and the output adapts to the aspect ratio of a supplied image.

### Can I combine reference material with first and last keyframes?

No. Reference and keyframe modes are mutually exclusive, so use one per request. Split the work across separate generations when you need both kinds of control.

### How do I hold a subject consistent across clips?

Pass source material through `inputReferences` and refer to each item by its order in the prompt, such as `Image 1`. MiniMax H3 conditions the generation on that reference, which keeps a character or product steady across shots.

### How do I call MiniMax H3 from the AI SDK?

Use `generateVideo`, imported as `experimental_generateVideo`, with the model set to `minimax/minimax-h3`. Pass your prompt plus optional `aspectRatio` and `duration`, and add `inputReferences` for reference-conditioned generations.

### Why do my requests time out before the video finishes?

Video generation takes longer than a text completion. Raise `pollTimeoutMs` in the provider options and extend fetch timeouts in Node.js so the request survives until the clip is ready.

### What does MiniMax H3 cost?

The listed video rate is N/A. Check the pricing panel on this page for the current rate and unit. AI Gateway reflects provider pricing with no markup and charges no platform fee on inference.

### Can I try MiniMax H3 before integrating it?

Yes. Use the playground on this page to generate clips and compare prompts, durations, and aspect ratios.

### Does MiniMax H3 support zero data retention?

Zero Data Retention is not currently available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

## Links

- [Model page](https://vercel.com/ai-gateway/models/minimax-h3)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://www.minimax.io/)