What MiniMax H3 Max is, how to call it, and why we built h3max.info

MiniMax H3 landed at the end of July 2026, the open weights followed in August, and fal.ai shipped a speed-tuned build called H3 Max a few weeks later. Since then every thread about it mixes the three names up. We kept re-explaining the difference, so we put the explanation on a site: h3max.info. This post is that explanation, plus the code to generate your first clip.

Three names, three different things

This is the part that trips everyone up.

The headline number is worth repeating: H3 Max renders a 5-second 768p clip in roughly three seconds, which is faster than the clip plays. That changes what you can build. People are already running live, chat-driven video streams on it.

Picture and sound in one pass

H3 is multimodal in both directions. One request takes text, an image, video or audio, and returns 5 to 15 seconds of video at 24 fps with a native stereo soundtrack. You are not stitching a silent clip to a separate audio model, and you are not fixing the audio afterwards either: it comes out of the same pass, so if you do not like it, you regenerate.

Calling it on fal.ai

The hosted endpoints are the quickest way in. Install @fal-ai/client, set your key, and subscribe:

import { fal } from "@fal-ai/client";

fal.config({ credentials: process.env.FAL_KEY });

const { data } = await fal.subscribe("minimax/h3-max/image-to-video", {
  input: {
    image_url: "https://example.com/portrait.jpg",
    prompt: "the person turns to camera and smiles, cinematic lighting",
    duration: 5,          // 5-15 seconds
    resolution: "768P",   // 480P or 768P
  },
});

console.log(data.video.url);

There are five endpoints. Text to video, image to video and reference to video on minimax/h3-max/…, plus text and image to video on minimax/h3-max-turbo/…. The inputs differ by mode:

• Text to video takes prompt, aspect_ratio, duration, resolution, seed.
• Image to video swaps the aspect ratio for image_url and an optional end_image_url to animate towards a final frame.
• Reference to video takes reference_image_urls, reference_video_urls and reference_audio_urls, which is how you keep one character consistent across shots.

subscribe blocks until the clip is done, which is fine in a script. In a web app, submit to the queue with a webhook instead and let the request return immediately — a 15-second render is still seconds of waiting you do not want to hold a connection open for.

What a clip actually costs

fal bills per second of generated video. The launch promotion ended on 7 September 2026, so these are the standard rates:

Two practical notes. Turbo at 480p is cheap enough to iterate on a prompt properly before you spend anything at 768p. And fal turns on auto top-up by default, so if you want a hard ceiling, switch it off in billing before you start experimenting.

Why we built h3max.info

The model is easy. Finding anything about it is not. The official announcement, the open weights, the API docs, the fal endpoints, the license, the deployment guides and the live projects are spread across half a dozen places, and search results are full of pages that confuse H3 with H3 Max.

So h3max.info started as two things, and has since grown a third:

• A directory — first the official H3 links, endpoints, weights and guides; since October 2026 a hand-written directory of AI tools across every category, from video generators to coding assistants. Every entry is read by a human before it goes in.
• A studio — the same model with the settings already decided, so you do not need an API key to try it.
• A guide — the MiniMax H3 Max page with what the model does, its options and what it costs.

Scenarios instead of settings

Every H3 Max playground asks the same questions: which endpoint, how long, what resolution, what aspect ratio, and then leaves you staring at an empty prompt box. Most people do not have an opinion on any of that. They have a photo and an idea.

So the Studio is built around scenarios. A scenario is one fixed model, one length, one resolution and a prompt written and tested in advance. What is left for you is a photo, a sentence, or both. The first one:

• Profile Page Takeover — upload an X profile screenshot and the person in the avatar climbs out of it and graffitis your line across the page. 15 s on H3 Max at 768p, with sound.

For everything else there is the AI video generator, where you pick H3 Max (or Hailuo) and write your own prompt. New accounts start with a few free credits, you see the price before you generate, and a render that fails refunds its credits automatically.

Things we ran into

• Files are big. A 15-second 1080p clip from H3 came back at 124 MB. Autoplaying a wall of those on a landing page is a quick way to burn your visitors' bandwidth, so we show poster frames and only load a clip when someone clicks it.
• fal hosts the output for you. The result URL is public, CORS-open and range-seekable, so you can point a <video> straight at it instead of copying the file. Retention is controlled per request, and files stay unless you set an expiry — but that also means the URL's lifetime is fal's call, not yours.
• The aspect ratio follows the input image on image-to-video, no matter what your product copy promises. Check what the model actually returned before you write "vertical" on the page.
• Prompt expansion is on your side. A hidden prefix that locks style, camera and mood does more for consistency than asking users to write better prompts.

Try it

If you want the API, the model APIs category lists fal, MiniMax and the other platforms that serve H3. If you just want a clip, open the Studio, pick a scenario and add a photo.

h3max.info is an independent community project. We are not affiliated with MiniMax or fal.ai, and prices and limits on their side change often — check the model page before you budget.