Images, video and audio together
Mix up to 9 images, 3 videos and 3 audio clips in one request: a character, a product, a camera move and a voice, each named in the prompt.
The MiniMax H3 API: MiniMax's Hailuo 03 turns a prompt, reference images, videos and audio into 768p or 2K video with sound, from $0.065 per second.
minimax-h3import os, time, requestsAPI = "https://api.zurelay.com/v1"headers = {"Authorization": f"Bearer {os.environ['ZURELAY_API_KEY']}"}res = requests.post(f"{API}/videos", headers=headers, json={ "model": "minimax-h3", "prompt": "A red fox trotting through fresh snow at dawn, cinematic", "seconds": 5, "resolution": "768p", "aspect_ratio": "16:9",})res.raise_for_status()job = res.json()# Takes a few minutes: poll until it's donewhile job["status"] in ("queued", "in_progress"): time.sleep(5) job = requests.get(f"{API}/videos/{job['id']}", headers=headers).json()if job["status"] == "failed": raise SystemExit(job["error"]["message"])video = requests.get(f"{API}/videos/{job['id']}/content", headers=headers)with open("video.mp4", "wb") as f: f.write(video.content)print(f"Saved video.mp4 for ${job['cost']}")Base URL https://api.zurelay.com/v1. Plain JSON over HTTPS: start a job, poll it, download the MP4. MiniMax H3 guide
Pricing
Pay as you go from prepaid credit, with no subscription. Top up from $10. A video is charged only once it's delivered; videos that fail cost nothing.
A clip costs its length in seconds times the price at its resolution. It's held from your balance when the job is accepted and charged when the video is delivered. Reference images, reference videos, reference audio and the aspect ratio don't change the price.
Every MiniMax model and pricePick a resolution and a length, and how many videos you make.
100 clips × $0.325
Overview
MiniMax H3 (Hailuo 03) is MiniMax's general video model, built to read text, images, video and audio together and follow detailed instructions about all of them. On Zurelay it makes 768p and 2K clips from 4 to 15 seconds long with native sound, from a prompt and up to 9 reference images, 3 reference videos and 3 reference audio clips in any mix. The Zurelay MiniMax H3 API starts at $0.065 per second, below MiniMax's own price, with the references included.
MiniMax released H3 in July 2026 as a general-purpose system for video: one model that takes text, images, video and audio as context and generates video with native stereo sound. It is also known as Hailuo 03. On Zurelay it uses the model ID minimax-h3, and also accepts MiniMax's own name, MiniMax-H3.
On Zurelay, MiniMax H3 makes video from a prompt and up to 9 reference images (people, products, places or a style), up to 3 reference videos whose motion and camera work it can follow, and up to 3 reference audio clips, as MP3 or WAV, for a voice or a soundtrack. Write the words in the prompt and a voice clip sets how they're spoken: in our tests the character said the lines word for word, in sync. Reference videos run 2 to 15 seconds each and 15 seconds in all, and so do audio clips. Zurelay doesn't offer H3 with a fixed first or last frame: give the image as a reference and say how the video opens.
Clips come in 768p or 2K, in whole seconds from 4 to 15, in 16:9, 9:16, 1:1, 4:3, 3:4 or 21:9. A 16:9 clip is 1376 × 768 at 768p and 1920 × 1088 at 2K, MiniMax's own sizes. Sound is always generated. The API is asynchronous: POST https://api.zurelay.com/v1/videos returns a job right away; poll GET /v1/videos/{id} until it's completed and download the MP4. A 5 second 768p clip usually takes about 2.5 minutes.
Zurelay prices MiniMax H3 by the second: $0.065 per second at 768p and $0.10 at 2K, against MiniMax's $0.08 and $0.13. MiniMax also bills reference video by the second and reference images past the fifth; on Zurelay, reference images, videos and audio cost nothing extra. The price is held from your balance when the job is accepted and charged only when the video is delivered; videos that fail cost nothing.
Strengths
Mix up to 9 images, 3 videos and 3 audio clips in one request: a character, a product, a camera move and a voice, each named in the prompt.
Write the words in the prompt and add a voice clip: the character says them in that voice, with the lips in sync.
2K comes out of the model in one pass at 1920 × 1088 for 16:9, not upscaled from 768p.
$0.065 per second at 768p and $0.10 at 2K, under MiniMax's own $0.08 and $0.13, and reference media never adds to the bill.
Use cases
A presenter image, a voice clip and the script in the prompt make a talking ad in 9:16, ready for social.
Give product photos as references and describe the scene, the camera and the sound.
Hand it a video whose movement or camera work you like and describe a new subject to carry it.
Add a music clip and ask for movement to the rhythm, for dances, loops and teasers.
Get started
No waitlist and nothing to install: three HTTPS endpoints and your API key.
Sign up, add credit and create an API key. Set a monthly budget or a rate limit per key if you like.
Every video endpoint lives under this base URL. Send your key as a Bearer token.
https://api.zurelay.com/v1POST a prompt to /v1/videos, poll the job every few seconds until it's completed, then download the MP4. Videos are kept for 30 days.
minimax-h3Works with the tools you already use
Compare
from $0.135 per second
Pick Seedance 2.5 for clips up to 30 seconds, up to 30 reference images, and editing or extending a video; it takes a first frame too.
Seedance 2.5 APIfrom $0.042 per second
Pick Gemini Omni Flash for a first and last frame, 360p to 4K, and video edits from a reference clip.
Gemini Omni Flash APIfrom $0.05 per second
Pick Wan 3.0 for a first frame, up to 10 reference images and clips up to 30 seconds.
Wan 3.0 APIZurelay charges by the second of video: $0.065 per second at 768p and $0.10 at 2K, so a 5 second 768p clip is $0.325. MiniMax's own price is $0.08 and $0.13 per second. Sound, reference images, reference videos, reference audio and the aspect ratio don't change the price.
Yes. Pass up to 3 MP3 or WAV clips in reference_audio, as https links or base64 data URLs, 2 to 15 seconds each and 15 seconds in all. A voice clip sets how a character sounds: write the exact words in the prompt, in quotes, and the character speaks them in that voice with matching lip movement. A music clip can be the soundtrack: say in the prompt to use the reference audio itself as the whole soundtrack, instrumental only, with no singing or text on screen, or the model may write its own song. Audio can come alone or with reference images and videos.
Not on Zurelay. MiniMax H3 here works from reference images instead: pass the image in reference_images and describe how the video opens. A request with first_frame or last_frame is refused at once with a message saying so, before anything is charged.
Send POST https://api.zurelay.com/v1/videos with your Zurelay key as a Bearer token and a JSON body: model set to minimax-h3, a prompt, and optionally reference_images, reference_videos, reference_audio, seconds (default 5), resolution (768p by default, or 2k) and aspect_ratio (default 16:9). You get a job back right away. Poll GET /v1/videos/{id} until its status is completed, then download the MP4 from /v1/videos/{id}/content.
A 5 second 768p clip usually takes about 2.5 minutes. 2K, longer clips and reference videos take longer. While you poll, the job's progress field gives an estimate from 0 to 100.
768p and 2K, in whole seconds from 4 to 15, in 16:9, 9:16, 1:1, 4:3, 3:4 or 21:9. A 16:9 clip is 1376 × 768 at 768p and 1920 × 1088 at 2K; MiniMax aligns frames to 32 pixels.
When Zurelay accepts the job, its price is held from your balance, and it's charged only when the video is delivered. If a video fails, including when the content filter refuses the prompt or a reference, you pay nothing.
Yes. Your requests run on MiniMax H3, and Zurelay checks that videos come back at H3's own frame sizes. Video generation varies from run to run, so the same prompt makes a different clip each time. Zurelay is an independent service and is not affiliated with MiniMax.
Create a key in seconds, point the SDK you already use at Zurelay, and every request costs up to 90% less from the first token.