Top AI Product

minWM is a full-stack open-source framework that walks you end-to-end through turning a bidirectional text-to-video foundation model into an action-conditioned video world model — the kind a game or agent can actually interact with frame by frame. It ships as a runnable tutorial, not just a paper.

## A pipeline you can stop at any step

The whole loop is open-sourced: data preparation, training, and inference, with input/output checkpoints exposed at every stage so you can pause, swap a component, or fork from any point. That’s the part most “open-source” world-model releases don’t ship — they publish weights, not the recipe. minWM bundles example data, runnable scripts, and Claude skills that capture hands-on experience and onboarding knowledge for newcomers.

## Why it matters

Real-time interactive video models — the technology behind playable AI-generated game worlds and visual simulators for embodied agents — have been gated by closed pipelines. Releasing the whole train-and-deploy stack lowers the floor: a researcher or studio can take a generic T2V model and turn it into something that responds to actions, without reverse-engineering anyone’s training process. As interactive video becomes a substrate for agents that need to “imagine” how the world will react, having an open, documented pipeline matters as much as the model that sits at the top of it.