Open model guide
H3-World: Keyboard-Controlled AI Video
H3-World turns MiniMax-H3 into something you drive rather than prompt. It is a community release, not an official product, and understanding that shapes what you should expect from it.
Use this workspace to turn a loose idea into a focused video brief with a subject, motion direction, reference style, and next scene cue.
H3-World is described as the first interactive world model built on MiniMax-H3.
Given an initial frame and keyboard state, it generates action-controlled video with coordinated character and camera motion.
It was published on Hugging Face by an independent researcher, not by fal or MiniMax.
How the controls work
You supply a starting image. Then, while the model generates, keyboard state is read and converted into a short language instruction for each upcoming block of video. The published description explains that each instruction is bound to its own stretch of frames so that a key press affects the right moment instead of bleeding across the whole clip.
What it is good at
The interesting result is coordinated motion: the character moves and the camera follows in a way that reads as intentional. For prototyping game-like interactions, testing how a scene responds to input, or building a small interactive demo, that is a real capability that plain text-to-video does not offer.
What it is not
H3-World does not build or store a scene. If you walk forward and then turn around, whatever is behind you is generated at that moment, shaped by what the model remembers, not loaded from a map. Long sessions can drift. Nothing can be exported as geometry. This is the same design as DeepMind's Genie 3 and Skywork's Matrix-Game, and it is a different category from World Labs' exportable worlds.
Running it
Because H3-World is an open release rather than a hosted endpoint, using it means running the model yourself. Check the Hugging Face page for weights, hardware expectations, and license terms before planning anything around it. Treat availability and quality as moving targets.
FAQ
Is H3-World the same as H3 Max camera controls?
No. fal's H3 Max Multi Angle sets a camera path from keyframes before generation. H3-World reads keyboard input during generation. One is directed, the other is driven, and neither produces a stored 3D scene.
Can I use H3-World for a live AI TV channel?
In principle, yes: it is built to keep generating in response to input. In practice you would need to host the model, handle input from viewers, and accept that a stream costs compute for every second it runs.
How does it compare to Genie 3?
Both are frame-by-frame interactive video models where consistency emerges from the model rather than from stored geometry. Genie 3 is not publicly available; H3-World is an open community release you can run yourself.