Header Logo

Docs

Start

Header Logo

Start

Header Logo

Docs

Start

All Models

MiniMax logo

MiniMax H3

MiniMax H3 generates cinematic video with native audio and dense reference control, including image, video, and audio inputs.

Price is per video output second, rounded up.
Offering
Price
480p
$0.01
768p
$0.036
2K
$0.13
Reference image (each) · 480p
$0.03
Reference image (first 5 free, then each) · 768p / 2K
$0.07
Reference audio (each) · 480p
$0.03
Reference audio · 768p / 2K
Free
Reference video (per input second) · 480p
$0.07
Reference video (per input second) · 768p
$0.13
Reference video (per input second) · 2K
$0.18

Call the API endpoint to get the latest, detailed pricing and pricing estimates.

Call the API endpoint to get the latest, detailed pricing and pricing estimates.

Call the API endpoint to get the latest, detailed pricing and pricing estimates.

MiniMax H3

MiniMax H3 is MiniMax’s open-weights video model, released in July 2026 from the company’s Shanghai lab. Open weights followed in August. It takes text, images, video, and audio in one request and returns a clip with native stereo sound.

You can generate from a prompt, from a first frame (and an optional last frame), or from a set of references. Reference-to-video accepts up to nine stills, up to three video clips with 15 seconds of video in total, and up to three audio tracks. An audio reference still needs something visual — a prompt, an image, or a clip — so the sound has a picture to attach to. Clips run from 4 to 15 seconds at 512p, 768p, or 1440p.

What it's good for

Shots that have to hold several things at once: a character that must stay on-model, a product that must stay recognizable, a camera move, and a line of dialogue or a score. H3 is built for the case where those constraints all matter in the same clip.

Model ID

minimax/minimax-h3

Output Type

Video

Pricing

From $0.01/s

Provider

MiniMax

Hosting

Protoface

MiniMax H3

MiniMax H3 is MiniMax’s open-weights video model, released in July 2026 from the company’s Shanghai lab. Open weights followed in August. It takes text, images, video, and audio in one request and returns a clip with native stereo sound.

You can generate from a prompt, from a first frame (and an optional last frame), or from a set of references. Reference-to-video accepts up to nine stills, up to three video clips with 15 seconds of video in total, and up to three audio tracks. An audio reference still needs something visual — a prompt, an image, or a clip — so the sound has a picture to attach to. Clips run from 4 to 15 seconds at 512p, 768p, or 1440p.

What it's good for

Shots that have to hold several things at once: a character that must stay on-model, a product that must stay recognizable, a camera move, and a line of dialogue or a score. H3 is built for the case where those constraints all matter in the same clip.