Skip to content
H3 Studio

Quickstart

Generate a clip with sound in three calls. You need an API key — create an account and issue one from the console.

1. Submit

curl https://api.h3.studio/v2/video_generation \
  -H "Authorization: Bearer $H3_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MiniMax-H3",
    "content": [{ "type": "text", "text": "a lighthouse at dusk, waves on black rock" }],
    "resolution": "768P",
    "duration": 5,
    "ratio": "16:9"
  }'

{ "task_id": "424010985738629" }

The call returns immediately. Generation takes 82–229 seconds depending on the tier, so nothing about this API blocks.

2. Poll

curl "https://api.h3.studio/v2/query/video_generation/$TASK_ID" \
  -H "Authorization: Bearer $H3_API_KEY"

{
  "task": {
    "id": "424010985738629",
    "model": "MiniMax-H3",
    "status": "succeeded",
    "created_at": 1785125529,
    "updated_at": 1785125946,
    "content": { "url": "https://cdn.h3.studio/outputs/...?Expires=…&Signature=…" },
    "resolution": "768P",
    "duration": 5,
    "usage": {
      "total_seconds": 5,
      "input_seconds": 0,
      "output_seconds": 5,
      "input_image_count": 0
    },
    "ratio": "16:9",
    "task_type": "generation",
    "modality": "video"
  }
}

status moves through queued → running → succeeded. Poll every few seconds, or skip polling entirely with a callback.

3. Download

content.url is a signed link that expires after an hour. It is minted fresh on every query, so if one expires, ask again.

curl -L "$URL" -o out.mp4

The result is an mp4: H.264 video at 24 fps, and an AAC stereo track at 32 kHz. The audio is generated jointly with the picture, in the same forward pass — it is not a soundtrack chosen afterwards.

In Python

import os, time, httpx

api = httpx.Client(
    base_url="https://api.h3.studio",
    headers={"Authorization": f"Bearer {os.environ['H3_API_KEY']}"},
)

task_id = api.post("/v2/video_generation", json={
    "model": "MiniMax-H3",
    "content": [{"type": "text", "text": "a lighthouse at dusk"}],
    "resolution": "768P",
    "duration": 5,
    "ratio": "16:9",
}).raise_for_status().json()["task_id"]

while True:
    task = api.get(f"/v2/query/video_generation/{task_id}").json()["task"]
    if task["status"] in ("succeeded", "failed", "cancelled"):
        break
    time.sleep(5)

if task["status"] != "succeeded":
    raise RuntimeError(task.get("error"))

with open("out.mp4", "wb") as f:
    f.write(httpx.get(task["content"]["url"], follow_redirects=True).content)

Writing a prompt that works

H3 was trained on structured prompts, and it shows. A prompt with explicit sections for picture, ambient sound and music produces markedly better results than a single sentence:

integrated_multimodal_description: [Shot 1] Cinematic, slow push-in. A lone
lighthouse stands on a rocky headland at dusk, its lamp rotating through thin
sea mist. Waves break against black rocks below.

overall_soundscape: The steady low roar of surf against rock. Wind over an
exposed headland. A gull calls twice, thin and far away.

non_diegetic_music: A slow, sparse piano figure with sustained low strings,
patient and unresolved.

MiniMax publishes a prompt-writing guide as an installable skill in the MiniMax-H3 repository; it applies here unchanged.

Next