MiniMax H3 vs H3 Max vs H3 Max Turbo: untangling the three video model names

One base model, three names

MiniMax released its H3 video model at the end of July 2026 and followed with open weights in August, offered under the MiniMax H3 Community License. A few weeks later, fal.ai shipped H3 Max, a build post-trained on H3 and tuned for speed, along with an even faster Turbo variant. According to a post on dev.to by the team behind the community site h3max.info, nearly every discussion about the model jumbles the three names together, which is why they wrote the explainer the rest of this article draws on.

The core distinctions, as laid out in that post:

• MiniMax H3 is MiniMax's own model. It reaches 2K resolution on the hosted API and 768p from the open weights, and it is the option for top quality, editing work or self-hosting.
• H3 Max is fal.ai's API-only build, capped at 768p. Its headline figure is roughly three seconds to render a five-second 768p clip, less time than the clip takes to play back.
• H3 Max Turbo is fal.ai's fastest, cheapest tier, also 768p and API-only, aimed at volume and iteration.

That sub-real-time speed is the significant number: the post reports people are already running live, chat-driven video streams on it.

Sound and picture in a single pass

H3 accepts text, images, video or audio as input and returns five to fifteen seconds of 24 fps video with a stereo soundtrack generated in the same pass. There is no step where a silent clip is married to a separate audio model; if the audio comes out wrong, the fix is regenerating the clip.

Five endpoints, five input shapes

On fal.ai, the hosted routes are split across the two variants. The h3-max namespace exposes text-to-video, image-to-video and reference-to-video; h3-max-turbo carries text-to-video and image-to-video.

Inputs differ by mode. Text-to-video takes a prompt, aspect ratio, duration, resolution and seed. Image-to-video replaces the aspect ratio with an image URL, plus an optional end image the clip animates towards. Reference-to-video accepts reference image, video and audio URLs, which is the mechanism for keeping a character consistent across shots.

One integration note from the post: the client's subscribe call blocks until the render finishes, which suits scripts but not web apps, where a queue submission with a webhook avoids holding a connection open through a fifteen-second render.

What it costs

fal bills per second of generated video, and a launch promotion ended on 7 September 2026. The standard rates listed in the post:

• H3 Max Turbo: $0.025 per second at 480p, $0.04 at 768p, which is $0.60 for a fifteen-second 768p clip.
• H3 Max: $0.05 per second at 480p, $0.08 at 768p, or $1.20 for the same clip.

The authors flag two practical points: Turbo at 480p is inexpensive enough to iterate on wording before spending anything at 768p, and fal turns on auto top-up by default, so anyone wanting a hard spending ceiling should switch it off before experimenting.

A community hub and a no-key studio

h3max.info itself has two parts. The hub is a hand-checked directory of official announcements, open weights, API docs, pricing, guides and community projects, with each entry reviewed by a person before inclusion. The studio wraps the same model in preset scenarios, a fixed endpoint, length, resolution, aspect ratio and pre-tested prompt template, so users supply only a photo and a sentence, with no API key required. Credits are prepaid and never expire, new accounts start with some for free, and failed renders are refunded. The project states it is independent of both MiniMax and fal.ai.

The build notes are a useful checklist for anyone shipping on this stack: output files are heavy, with a fifteen-second 1080p H3 clip weighing in at 124 MB, so the site shows poster frames and loads video on click; fal hosts outputs at public, CORS-open, range-seekable URLs whose retention is set per request; image-to-video inherits the input image's aspect ratio regardless of what a product page claims; and a hidden prompt prefix pinning style, camera and mood does more for consistency than asking users to write better prompts.

Why it matters

Generation faster than playback changes the category a video model can serve: batch content production becomes an interactive medium, and the live chat-driven streams already running on H3 Max are the early evidence. The split between open weights on H3 and API-only tuned builds also gives teams a real trade-off between control and speed. And the naming confusion the post set out to fix is itself a practical hazard: pricing, resolution ceilings and licensing differ across the three names, so knowing exactly which one a benchmark, tutorial or pricing page refers to determines what you can build and what it will cost. Since all specifics here come from a single community post with ties to h3max.info, the post's own closing advice applies: check the model page for current prices and limits before budgeting.