Video

MiniMax H3

⧉

minimax-h3

MiniMax H3（Hailuo 03）视频模型：1440P 高清输出，固定生成原生音频，时长 5–15 秒逐秒可选，六种画幅。不带参考图为文生视频、带参考图（作首帧）为图生视频。异步任务：提交后轮询任务号获取重托管后的稳定 mp4 直链，按秒计费、失败全额退。

文生视频图生视频1440P原生音频

Get API Key → API Docs

Live Test · Playground

Try out MiniMax H3 right here (available after login)。

Input

Video

Prompt

0 / 2000 bytes

Duration

Resolution

Aspect ratio

Reference image role

Reference image URLs (optional, one per line, up to 5)

With multiple references, address them in the prompt as @Image1 / @Image2, e.g. "the character from @Image1 stands in the scene from @Image2". Without tokens the model decides on its own.

Reference audio URLs (optional, one per line, requires a reference image; up to 3)

Rendering takes 2-12 min (longer clips take longer); billed per second, fully refunded on failure. Prompt limit 2000 bytes (about 666 Chinese characters).

Output

video

Results will appear here after running.

About MiniMax H3

MiniMax H3 (Hailuo 03) is the highest-resolution video model here: 1440P output with native audio always on. Duration is any integer 5-15 seconds across six aspect ratios; omit the reference image for text-to-video, include one (first frame by default) for image-to-video. Async job: POST /v1/videos/generations returns HTTP 202 and a job id immediately; poll GET /v1/videos/jobs/{id} until status=succeeded and read data[0].url. That link is an mp4 re-hosted by us — it will not expire the way the upstream CDN link does.

Use cases

High-definition output

2560x1440, with the best per-frame detail and lighting of any video model on the station.

Clips with sound, no audio pass

Native audio is always generated with no extra parameters, removing a whole production step.

Full camera choreography

It handled a tracking shot into a whip-tilt into a low-angle finish, so it suits work with a real shot list.

Image-to-video

With the reference as the first frame, subject and composition stay stable — good for turning concept art into motion.

How to choose

Pick it when you want resolution, native audio, or full camera choreography. For more than 15 seconds use seedance-2.5; to turn the soundtrack off, or to drive from a source video, use the Seedance line. It also renders the slowest here (about 11 minutes for a 15-second clip), so use seedance-2.0-fast when speed matters.

FAQ

Can I turn the audio off?

No. This model always generates native audio, and audio=false is explicitly rejected rather than silently ignored. For silent clips use the Seedance line with audio=false.

Why does it take so long?

1440P with native audio is simply heavy: roughly 4 minutes for 5 seconds and 11 minutes for 15. It is an async job, so you hold a job id and poll — no connection is kept open.

How does image-to-video work?

Pass an image (public URL, data: URI or base64). By default it becomes the first frame, so the clip keeps the reference subject and framing. A second image becomes the last frame; set image_role=reference to switch to subject/style reference mode instead. The first/last-frame and reference modes are mutually exclusive.

How is it billed, and am I charged on failure?

Billed per second: the hold placed at submit is computed from the duration you asked for, and the settle is the exact same number. A failed or timed-out render refunds the whole hold, so you never pay for a clip you did not receive.

Related Models

Explore other models you can integrate.

MiniMax H3

Live Test · Playground

Input

Output

About MiniMax H3

Use cases

How to choose

FAQ

Related Models

GPT-5.6 Sol

GPT-5.6 Terra

GPT-5.6 Luna

GPT-5.5

GPT Image 2

Nano Banana 2

Nano Banana Pro

Claude Sonnet 4.6

Claude Opus 5

Claude Fable 5

Gemini 3.1 Pro

Gemini 3.6 Flash

Gemini 3.6 Flash High

Gemini 3.6 Flash Low

Gemini 3.6 Flash Tiered

Gemini 3 Flash

Gemini 3.5 Flash

Gemini 2.5 Flash

Veo 3.1

Seedance 2.5

Seedance 2.0

Seedance 2.0 Fast

Welcome back

Create account

Reset password

Contact the Author