MiniMax H3 Max Turbo

Overview

MiniMax H3 Max Turbo is the speed tier of H3 Max, built by fal. MiniMax open-sourced H3 in August 2026, fal post-trained those weights into H3 Max for stronger prompt adherence and better aesthetics, and then distilled H3 Max again for throughput. Turbo is that third step, released as a preview in September 2026. What comes back is a 5 to 15 second clip at up to 768p and 24fps, with stereo audio generated alongside the picture.

The claim, and what it rests on

fal puts Turbo at twice the speed of H3 Max, targeting the 97th percentile of H3 Max quality on their own evaluations, and reports that it still scores above the base MiniMax H3 it descends from. That last clause is the one that matters. A fast tier normally sits below the model it was cut down from, and the interesting thing here is where the floor landed: below H3 Max, above base H3.

The number worth holding on to is not the multiplier but what it crosses. H3 Max already returned a five-second clip in under three seconds. Doubling that puts a fifteen-second generation comfortably inside fifteen seconds of waiting, which means the render finishes before the clip could have played. Video generation stops being something you wait for and starts being something that keeps up with you.

What the speed changes

At half a minute, a generation is a submission you walk away from. At a second and a half, it is closer to a keystroke. Twenty variations of an ad hook get judged as finished motion instead of as a storyboard. A shot gets blocked, watched, adjusted, and watched again inside a single sitting. Work that was never worth the wait, exploring a direction you only half believe in, becomes cheap enough to simply try.

It also opens a category that was closed at thirty-second latency: anything with a person waiting on the other end. A generate button inside a product, a chat-driven loop, a demo running live in front of an audience. None of those tolerate a spinner measured in minutes.

Where it is the wrong choice

Two places, and they are easy to name. Reference to video, the mode that holds a character or a style steady across shots from supplied images, clips, and audio, is not on Turbo. That belongs to H3 Max. And the 768p ceiling is inherited, not chosen: H3 ships as three parts, a preprocessor that interprets the brief, a base model that renders at 768p, and a third stage that regenerates the result at 2K using the original brief so fine detail and small text are recovered rather than guessed at. MiniMax open-sourced only the middle part, so nothing derived from those weights has a 2K stage behind it.

That makes the pairing obvious rather than clever. Explore on Turbo, finish on base H3. Both come from the same weights, so the look carries across the handoff, which is not true when you draft on one vendor's model and finish on another's.

MiniMax H3 Max Turbo on Morphic

MiniMax H3 Max Turbo sits in the video model picker alongside MiniMax H3 Max, MiniMax H3, Seedance, Kling, and Veo. Write the shot as a brief, generate, and the take lands on the Canvas, where you can run the same brief through another model and compare them side by side. The habits this model rewards, writing the action across the clip rather than describing a frame, and directing sound as deliberately as picture, carry to every one of them.