MiniMax H3

Request shape

A generation needs a positivePrompt. Everything else is optional: width and height pick the format, duration sets the length, and seed makes a result repeatable.

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'minimax:h3@0',
  positivePrompt: 'A golden-hour beauty and fashion shot. A model with long windswept hair stands in a field of tall golden grass, turning slowly toward camera as warm sunlight flares through the lens, her hair and dress flowing in the breeze. Slow cinematic push-in, shallow depth of field, photoreal. Sound: a gentle steady wind, dry grass rustling, a warm ambient tone.',
  width: 2560,
  height: 1440,
  duration: 8
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "minimax:h3@0",
            "positivePrompt": "A golden-hour beauty and fashion shot. A model with long windswept hair stands in a field of tall golden grass, turning slowly toward camera as warm sunlight flares through the lens, her hair and dress flowing in the breeze. Slow cinematic push-in, shallow depth of field, photoreal. Sound: a gentle steady wind, dry grass rustling, a warm ambient tone.",
            "width": 2560,
            "height": 1440,
            "duration": 8
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "videoInference",
      "taskUUID": "b7c8d9e0-1f2a-4b3c-8d4e-5f6a7b8c9d0e",
      "model": "minimax:h3@0",
      "positivePrompt": "A golden-hour beauty and fashion shot. A model with long windswept hair stands in a field of tall golden grass, turning slowly toward camera as warm sunlight flares through the lens, her hair and dress flowing in the breeze. Slow cinematic push-in, shallow depth of field, photoreal. Sound: a gentle steady wind, dry grass rustling, a warm ambient tone.",
      "width": 2560,
      "height": 1440,
      "duration": 8
    }
  ]'

runware run minimax:h3@0 \
  positivePrompt="A golden-hour beauty and fashion shot. A model with long windswept hair stands in a field of tall golden grass, turning slowly toward camera as warm sunlight flares through the lens, her hair and dress flowing in the breeze. Slow cinematic push-in, shallow depth of field, photoreal. Sound: a gentle steady wind, dry grass rustling, a warm ambient tone." \
  width=2560 \
  height=1440 \
  duration=8

{
  "taskType": "videoInference",
  "taskUUID": "b7c8d9e0-1f2a-4b3c-8d4e-5f6a7b8c9d0e",
  "model": "minimax:h3@0",
  "positivePrompt": "A golden-hour beauty and fashion shot. A model with long windswept hair stands in a field of tall golden grass, turning slowly toward camera as warm sunlight flares through the lens, her hair and dress flowing in the breeze. Slow cinematic push-in, shallow depth of field, photoreal. Sound: a gentle steady wind, dry grass rustling, a warm ambient tone.",
  "width": 2560,
  "height": 1440,
  "duration": 8
}

[
  {
    "taskType": "videoInference",
    "taskUUID": "b7c8d9e0-1f2a-4b3c-8d4e-5f6a7b8c9d0e",
    "videoUUID": "3a4b5c6d-7e8f-4a0b-9c1d-2e3f4a5b6c7d",
    "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/3a4b5c6d-7e8f-4a0b-9c1d-2e3f4a5b6c7d.mp4"
  }
]

The prompt runs up to 2,000 characters, which is plenty for a detailed shot with its soundscape. Leaving width and height out lets the model pick a format. The next section covers when to set them.

Sizes and aspect ratios

H3 renders at 2K, and only at a fixed set of aspect ratios. Pick one by setting the exact width and height pair, or omit both and let the model choose:

• 16:9 at 2560 × 1440 (landscape, the default feel)
• 21:9 at 2944 × 1248 (cinemascope)
• 4:3 at 1920 × 1440
• 1:1 at 1440 × 1440
• 3:4 at 1440 × 1920
• 9:16 at 1440 × 2560 (vertical)

There is no free sizing and no other resolution tier: set one of these pairs or neither. Passing a width and height that do not match a pair is rejected, and passing only one of the two is rejected. The three below are one prompt at three formats, and the ratio is not a crop, it changes what the shot holds:

The 16:9 frame holds the walk and the architecture around her. The vertical 9:16 crops in close and stacks the tall windows overhead, which is the shape for social. The 21:9 cinemascope opens the sides and gives the atrium its widescreen scale. Choose the format before you write, because "walks through the atrium" reads differently when the frame is tall than when it is wide.

Duration

duration is an integer number of seconds from 5 to 15, defaulting to 5. Longer is not automatically better: a 15-second clip is one continuous take, so it needs a subject with enough sustained motion to fill it. The potter below earns the full length because the action keeps evolving across the whole shot:

For a punchy beat, a jump, a reveal, stay at 5 to 8 seconds. Match the length to how much genuinely happens, since padding a simple action out to 15 seconds tends to stall or loop rather than develop.

Native audio, and prompting for it

Audio is the feature to build around. Every clip comes back with synced stereo sound generated alongside the picture, and the model takes direction on it the same way it takes direction on the image: describe the soundscape in the prompt and it renders that. Write a Sound: clause naming the effects, ambience, and timing you want.

The prompt asked for the ring of the hammer timed to the moment of impact, and the model syncs the sound to the action rather than laying a generic loop over the top. That timing is the trick: name the sound and when it happens ("the hammer rings out the instant it strikes", "steam hisses as the blade is quenched"). Ambience works the same way, so "a calm ambient hum and soft footsteps" fills the atrium without any on-screen source.

Directing the shot

H3 reads cinematic language: name the camera move, the lens feel, the lighting, and the pace, not just the subject. The shot below is built from shot-craft terms rather than a scene description:

"Low-angle dolly-in", "chiaroscuro lighting", and "raking across the bodywork" each did visible work. This is where the model is strongest: atmospheric, motion-led, single-shot footage with a clear camera idea and a matching soundscape. It is weaker at things a single prompt cannot pin down, like exact on-screen text, precise object counts, or holding one specific character's face across a whole clip. Those are jobs for reference inputs rather than a text prompt.

Tips

1. Write the sound, not just the picture. A Sound: clause naming effects and ambience is the difference between H3's audio working for you and the model guessing.

2. Time key sounds to on-screen events. "The hammer rings as it strikes" syncs, while "forge sounds" drifts. The same holds for impacts and footsteps.

3. Pick a format from the six pairs. Set an exact width and height pair or omit both. Any other size is rejected.

4. Match duration to the action. Reach for 15 seconds only when the motion keeps developing. Hold short beats at 5 to 8.

5. Direct like a cinematographer. Name the camera move, the lens, and the lighting. "Slow low-angle dolly-in, chiaroscuro" beats "a cool shot of a car".

6. Lock a seed once a clip lands. Video is slow to iterate, so a fixed seed keeps a result you like while you adjust the wording around it.