2K by default
MiniMax offers 2K as the standard resolution, made possible by a new video tokenizer it calls H3-VAE. You can also pick 768p.
MiniMax's newest video model, released in July 2026. It makes the picture and stereo sound together, from a text prompt or from your own images, at up to 2K.
On this page you can make 5-, 10- or 15-second clips at 768p or 2K, from text, from a start and end frame, or from reference images, videos and audio. Sound is always on.
MiniMax H3 is a video generation model from MiniMax, released on July 31, 2026, and the successor to its Hailuo 01 and Hailuo 02 models. MiniMax describes it as a general-purpose multimodal model: it reads text, images, video and sound as one context and generates video with native stereo audio, up to 15 seconds at 2K.
Instead of separate models for text to video, image to video, start and end frames or references, MiniMax trained one model for all of them, with the relationship between your inputs described in plain language. This page offers all three: text to video, start and end frames, and references (up to 9 images, 3 videos and 3 audio clips).
From the MiniMax H3 announcement and the model documentation.
MiniMax offers 2K as the standard resolution, made possible by a new video tokenizer it calls H3-VAE. You can also pick 768p.
Voices, sound effects and music are modeled together with the video and come out in native stereo, with no separate audio step.
MiniMax says early testers found it strong at following instructions and at showing text and brand details accurately, which is why it pitches the model for ads, product pages, e-commerce and game UI.
What follows from the model documentation and MiniMax’s own examples.
Say who moves, how, and what the camera does: “the camera slowly pushes in”, “she turns toward the window”. A prompt that only describes a scene tends to give a clip where little happens.
This model always generates sound. Name what we should hear (a line of dialogue in quotes, rain on a tin roof, soft piano), otherwise it picks something for you.
MiniMax says H3 models several shots natively. To get cuts, describe them in order (“Shot 1… Shot 2…”); for one continuous take, say so.
With references, say what each one is for and call it by the name shown under it, the way MiniMax’s own examples do: “the person in Image 1 sings, with the camera move from Video 1 and the voice from Audio 1”.
The first image is the opening frame, so the prompt only needs the movement from there; with an end frame, describe how to get from one to the other. The video keeps the first image’s shape.
Yes. Every video on this page is generated with MiniMax H3, not another model.
Videos are charged per second, and 2K costs more per second than 768p. The total for the length and quality you pick is shown on the button before you generate, and new accounts receive free credits.
Much longer than an image, and it varies with demand. You can leave the page: the video keeps generating and shows up in My tasks when it is ready.
No. MiniMax H3 always generates sound together with the picture. If you need a silent clip, mute it when you edit.
Without an image, the model needs a fixed shape, so pick one (16:9 by default). When you start from an image, the video keeps that image’s shape.
Reference audio is free. Reference videos are charged per second at the same rate as the video you generate, and images beyond the fifth cost a little extra. The button shows the total before you generate.
Up to 15 seconds, which is also the longest the model itself goes according to MiniMax. Choose 5, 10 or 15 seconds at 768p or 2K; the price is per second.
Yes, as an MP4 file. Results are kept for 30 days, so download anything you want to keep.
Prompts, images and results are checked for sexual and other sensitive content, and flagged requests are declined. Rephrase the prompt or use a different image and try again. Credits for declined or failed requests are refunded.