Stable WorldModel is a new open-source platform from the galilai-group, an academic consortium that includes Yann LeCun and Randall Balestriero among its authors, designed to make world-model research reproducible end-to-end. World models are arguably the most contested area in modern AI: a single repo that competently handles data collection, training, and model-predictive control evaluation across 25+ environments is exactly what the field has been missing. By the end of May 2026 the project has crossed 1,300 GitHub stars off the back of a release-quality v0.1.0 and an arXiv paper (2605.21800). ## Why World Models Need a Stable Substrate Researchers in this space have spent disproportionate energy on plumbing rather than ideas: writing dataset loaders, hooking up environments, re-implementing baselines, and gluing together planners. Stable WorldModel collapses all of that into a single library so that the interesting variable is the model under study. Reference implementations are provided for DINO-WM, LeWM, PLDM, GCBC, GCIVL, and GCIQL, alongside classical planning solvers like CEM, iCEM, MPPI, Predictive Sampling, gradient methods, and an Augmented Lagrangian for constrained settings. ## A Format Registry, Not a Format Lock-In One of the most pragmatic decisions in the project is the data format registry. Datasets can be recorded, loaded, or converted across LanceDB (default, append-friendly, column-oriented), HDF5 (single-file portable), folder format with NPZ plus JPEGs (inspection-friendly), MP4-backed video format (compact storage), and a read-only LeRobot adapter for Hugging Face datasets. Throughput benchmarks reported in the README show LanceDB at 4,814 samples/second locally versus HDF5 at 1,416, and 3,183 over S3 versus 9.1 for HDF5 without caching, while cutting on-disk size from 43 GB to 13 GB on a representative dataset. ## 25+ Environments Out of the Box The environment suite spans DeepMind Control Suite, Gymnasium classics, OGBench cube and scene manipulation, Fetch robotics tasks, Craftax pixel and symbolic variants, the full Atari Arcade Learning Environment with over 100 games, and dedicated world-model benchmarks like PushT and TwoRoom. Crucially, most environments expose visual and physical factors of variation (lighting, textures, morphology, dynamics) as first-class controllable parameters, so zero-shot distribution-shift evaluation does not require building a custom rig. ## CLI for Everyday Research The swm CLI lets researchers list cached datasets, inspect them, browse environments and their variation parameters, list checkpoints, and convert datasets between formats without writing code. This is the kind of detail that separates a research drop from a research platform: the small ergonomic wins that compound across a research group of a dozen people. ## Authors and Provenance The author list includes Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, and Randall Balestriero, among others. The associated arXiv preprint (2605.21800v1) lays out the motivation: world models will succeed or fail on reproducibility, and a shared infrastructure layer is now table stakes. Stable WorldModel is positioned as that layer. ## Practical Notes Stable WorldModel installs from PyPI with optional [all] extras for training, environments, and data format support. Datasets and checkpoints live under STABLEWM_HOME (default ~/.stable_worldmodel/). LeRobot support requires Python 3.12+. The project is Python 99.7% by line count, with a clean release at v0.1.0 (May 26, 2026). ## Limitations As a v0.1.0 release, APIs may change between minor versions, and the authors say so explicitly. Documentation is good for the core path but thinner for edge cases like writing a brand-new environment family from scratch. The included baselines are competent reference implementations, not state-of-the-art tuned models, which is the right choice for a platform but means head-to-head benchmark numbers should be reproduced rather than cited. Finally, world-model planning at scale is still GPU-hungry and storage-hungry; LanceDB helps with the storage side, but compute budgets remain a real constraint.