A common starting point
World-model research spans representation learning, predictive dynamics, planning, and the environments in which ideas are tested. Many models need the same basic components, even when the ideas behind them differ.
PyTorch World brings those shared parts into one library. Layers, masking, objectives, planners, and environments can be reused, making a new model a shorter file and existing models easier to compare.
From representations to actions
The library includes joint-embedding architectures for images, video, and vision-language tasks, alongside action-conditioned world models. A shared interface connects prediction, evaluation, and planning.
Examples and notebooks provide a practical path into the code. Researchers can explore pixel environments or MuJoCo tasks, try a model, and work toward their own implementation.
Open foundations
PyTorch World is released under the Apache 2.0 license. The core depends only on PyTorch, with optional dependencies for examples and action environments.
uv add torch_worldThis is where 
Latent Horizon’s work begins: useful tools, clear implementations, and a foundation others can build on.