MiniMax H3 shipped with open weights and a claim that sounded too broad to be real: one model reading text, images, video, and audio in a single context, returning 2K clips with stereo sound already mixed to picture.
I spent the last day trying to find the least annoying way to actually run it, and fal is where I stopped looking.
TL;DR
- fal is the best place to run MiniMax H3 because it hosts the open weights on its own serving infrastructure as a Day 0 partner, so all three endpoints have been callable since launch with no GPUs to rent, no images to build, and no waitlist to sit through.
- Every generation returns 5 to 15 seconds at 24 FPS in 2K with native stereo audio, and prompts can run up to 7,000 characters.
- Output costs $0.26 per second, which puts a 5-second clip at $1.30 and a full 15-second one at $3.90, billed per use with no subscription and no minimum spend.
Where can you access MiniMax H3 right now?
fal is where I’d send anyone who wants H3 working this afternoon.
There are three endpoints, and they map onto the things you’d actually want to do without much overlap.
- minimax/h3/text-to-video is prompt only, and the input schema is short enough to memorize: prompt, duration (defaults to 5), resolution (2K is currently the only accepted value), and aspect_ratio, which defaults to 16:9 with 21:9, 4:3, 1:1, 3:4, and 9:16 as the alternatives.
- minimax/h3/image-to-video handles first and last frame work, so you pass an opening image, optionally a closing one, and H3 fills the motion between them while output follows the aspect ratio of whatever you uploaded.
- minimax/h3/reference-to-video is the one I found hardest to put down, accepting up to 9 reference images, 3 video clips of 2 to 15 seconds each, and 3 audio tracks, capped at 12 files total in a single request.
Every endpoint returns a single video file with a URL, content type, file name, and byte size, so piping results into storage or a render queue takes no parsing gymnastics.
There’s also a browser playground on each model page, which is honestly where I did most of my prompt fiddling before writing a line of code.
Here’s how an API call looks with their image-to-video endpoint:
import { fal } from “@fal-ai/client”;
const result = await fal.subscribe(“minimax/h3/image-to-video”, {
input: {
prompt: “The camera slowly pulls back from the scene, revealing the full landscape as clouds drift overhead and light shifts across the terrain.”,
image_url: “https://storage.googleapis.com/falserverless/example_inputs/hailuo23/pro_i2v_in.jpg”
},
logs: true,
onQueueUpdate: (update) => {
if (update.status === “IN_PROGRESS”) {
update.logs.map((log) => log.message).forEach(console.log);
}
},
});
console.log(result.data);
console.log(result.requestId);
I also found that, for longer jobs, fal.queue.submit with a webhook URL stops you from blocking on a 15-second 2K render.
What makes fal the best place to use MiniMax H3
fal is the best place to use MiniMax H3 because it absorbs every piece of infrastructure the model needs and hands you back a single function call.
The open weights are what make that possible, and not for the reason most posts cite.
As they’re public, fal isn’t proxying somebody else’s hosted API and marking it up.
The team has the model itself and runs it on their own serving stack, and that’s where the difference between a wrapper and a real host shows up.
You get the practical benefit without the ops bill: no GPUs to rent, no container images to babysit, no cold starts to reason about.
And the escape hatch stays open, since the weights are still yours to pull down and fine-tune if a project ever justifies it.
Commercial use is covered on the API too, which saved me a licensing conversation.
How much does it cost to run MiniMax H3 on fal?
Running MiniMax H3 on fal costs $0.26 for every second of finished video at 2K.
Since 2K is the only resolution H3 currently outputs, duration is your one cost lever, which makes budgeting refreshingly boring.
Five seconds runs $1.30, ten seconds runs $2.60, and the maximum 15 seconds runs $3.90.
There’s no subscription, no seat fee, and no minimum, so a week of poking around at 5-second clips costs less than a lunch.
My advice: lock your prompt at 5 seconds, then extend once the shot is behaving.
Run MiniMax H3 on fal
I went looking for the shortest path between an idea and a 2K clip with sound, and this was it.
You can start with minimax/h3/text-to-video to calibrate how H3 reads a prompt, because it responds to camera and audio direction more literally than I expected.
Once you have assets worth bringing, you can move to reference-to-video and let it carry identity, motion, and voice through in one pass.
