Inference Logoinference shell
apps/minimax/h3

h3

MiniMax H3 — multimodal video generation with native audio. Text-to-video, image-to-video with end_image, and reference-based generation. 2K resolution, 5-15s duration, 24fps. Prompt with timelines, audio cues, and negative lists.

run in browser →run via API
run with your agent
# install belt
$curl -fsSL https://cli.inference.sh | sh
# view schema & details
$belt app get minimax/h3
# run
$belt app run minimax/h3

api reference

about

minimax h3 — multimodal video generation with native audio. text-to-video, image-to-video with end_image, and reference-based generation. 2k resolution, 5-15s duration, 24fps. prompt with timelines, audio cues, and negative lists.

1. calling the api

install the client

the client provides a convenient way to interact with the api.

bash
1pip install inferencesh

setup your api key

set INFERENCE_API_KEY as an environment variable. get your key from settings → api keys.

bash
1export INFERENCE_API_KEY="inf_your_key"

run and get result

submit a request and wait for the final result. best for batch processing or when you don't need progress updates.

python
1from inferencesh import inference23client = inference()456result = client.run({7        "app": "minimax/h3",8        "input": {}9    })1011print(result["output"])

stream live updates

get real-time progress updates as the task runs. ideal for showing progress bars, partial results, or long-running tasks.

python
1from inferencesh import inference23client = inference()456# stream=True yields updates as they arrive7for update in client.run({8        "app": "minimax/h3",9        "input": {}10    }, stream=True):11    if update.get("progress"):12        print(f"progress: {update['progress']}%")13    if update.get("output"):14        print(f"output: {update['output']}")

2. authentication

the api uses api keys for authentication. see the authentication docs for detailed setup instructions.

3. files

file inputs are automatically handled by the sdk. you can pass local paths, urls, or base64 data.

automatic upload

the python sdk automatically detects local file paths and uploads them. urls are passed through as-is.

python
1# local file paths are automatically uploaded2result = client.run({3    "app": "minimax/h3",4    "input": {5        "image": "/path/to/local/image.png",  # detected & uploaded6        "audio": "https://example.com/audio.mp3",  # url passed through7    }8})

manual upload

you can also upload files manually and use the returned url.

python
1# upload and get a hosted URL2file = client.files.upload("/path/to/file.png")3print(file.uri)  # https://cloud.inference.sh/...

4. webhooks

get notified when a task completes by providing a webhook url. when the task reaches a terminal state (completed, failed, or cancelled), a POST request is sent to your url with the task result.

python
1result = client.run({2    "app": "minimax/h3",3    "input": {},4    "webhook": "https://your-server.com/webhook"5}, wait=False)

webhook payload

your endpoint receives a JSON POST with the task result:

json
1{2  "id": "task_abc123",3  "status": 9,4  "output": { ... },5  "error": "",6  "session_id": null,7  "created_at": "2024-01-15T10:30:00Z",8  "updated_at": "2024-01-15T10:30:05Z"9}
idstring— task id
statusnumber— terminal status (9=completed, 10=failed, 11=cancelled)
outputobject— task output (when completed)
errorstring— error message (when failed)
session_idstring— session id (if using sessions)
created_atstring— iso timestamp
updated_atstring— iso timestamp

5. schema

input

promptstring*

video prompt. use six-block structure: style contract, timeline with [0s-2s] markers, camera, audio cues, spelled-out text, negative list. max 7000 chars.

example: "[0s-3s] Static wide shot. A lone marble camera on a pedestal in darkness, single moonbeam. [3s-6s] The lens flares to life, film strips unfurl from the reels. [6s-8s] Film strips fill the frame. Camera: locked off, no cuts. Audio: low hum building to orchestral swell at 6s. Do not add subtitles."
imagestring(file)

first frame image for image-to-video. formats: jpg, png, webp, heic. max 30mb.

last_imagestring(file)

last frame image. requires image to be set as first frame.

reference_imagesarray

reference images for style/content guidance. cannot combine with first/last frame.

reference_videostring(file)

reference video for motion/style guidance. mp4/mov, max 50mb, 2-15s.

reference_audiostring(file)

reference audio for synchronized generation. wav/mp3, max 15mb, 2-15s.

resolutionstring

output resolution. 2k is the only resolution currently accepted.

default: "2K"
options:"768P""2K"
durationinteger

video duration in seconds. 5-10 for text/image-to-video, 5-15 for reference-to-video.

default: 8min:5max:15
ratiostring

aspect ratio. text-to-video requires explicit value (not adaptive). image/reference modes default to adaptive.

default: "16:9"
options:"adaptive""21:9""16:9""4:3""1:1""3:4""9:16"

output

videostring(file)*

the generated video file.

ready to run h3?

try in browserbrowse all tools

we use cookies

we use cookies to ensure you get the best experience on our website. for more information on how we use cookies, please see our cookie policy.

by clicking "accept", you agree to our use of cookies.
learn more.