Nano World Model is a minimalist, out-of-the-box codebase whose goal is to advance World Model scientific research. It trains video world models based on the diffusion-forcing method, and its overall design emphasizes minimal dependencies and easy data loading, claiming that going from cloning the repository to the first rollout takes only a few minutes. In May 2026, the project released a technical report (arxiv), all model checkpoints (Hugging Face), and training, inference, and evaluation code.
In terms of main capabilities, the repository provides a unified pipeline covering training, validation, and evaluation, all managed through a clear Hydra configuration system. It emphasizes scientific transparency, has a clean codebase, supports directly comparable ablation experiments across prediction targets, action injection, and model scale, and is fully open source, including model checkpoints. Its application scenarios are diverse, supporting long-horizon rollout, converting rollouts into 3D point clouds, and MPC (Model Predictive Control)-based planning out of the box.
In terms of usage, you can first try the Colab quick start notebook, download the official Point Maze checkpoint, and preview prediction results side by side with recorded validation frames; after selecting a GPU runtime, run all cells, and the notebook uses the same locked configuration as below. uv must be installed first. The supported Linux x86_64 environment uses a single commit lock to manage Python, PyTorch, CUDA wheels, and all dependencies; see the relevant documentation for environment setup and verification. For the locked RT-1/LeRobot dependencies, run ./nanowm sync --extra rt1 ; note that lerobot-datasets should not be installed, as it is a dataset format version rather than a PyPI package. Then you can set the data and results paths, or use the git-ignored src/configs/local/paths.yaml template. The i3d torchscript required for FVD evaluation needs to be downloaded, and the InceptionV3 used for FID will download itself on first use. For how to download the datasets (DINO-WM, RT-1, CSGO), see docs/datasets/README.md .
Training examples include: DINO-WM PushT uses NanoWM-B/2 with default settings (pred-v, additive injectioN, cosine + ZTSNR); CSGO uses the L/2 model; RT-1 (fractal) is run as the primary. For reproducibility, example scripts are provided in src/scripts/, and the full training guide, design choices, and ablation table are in docs/training.md. The repository also provides pretrained checkpoints for the best configuration runs, along with the RT-1 ablation table and the corresponding Hugging Face checkpoint paths.
For sample predictions, it includes a CSGO 50-frame autoregressive long rollout (NanoWM-L/2, 100k). Quantitative metrics are evaluated on 256 fixed samples (seed=42), 250-step DDIM, sequential scheduling (frame-by-frame autoregressive denoising); the complete numbers by domain and methodology are in docs/evaluation.md. NanoWM rollouts can be directly used for downstream applications, including long-horizon generation, video-to-3D reconstruction, and MPC-style planning: long-horizon rollouts are generated autoregressively from training checkpoints; video-to-3D map uses Depth Anything 3 point cloud reconstruction; MPC-style planning performs CEM planning on world model rollouts.
The documentation directory includes: docs/config_system.md (Hydra configuration layout, overrides, environment variables), docs/training.md (training workflow, design choices, ablation table, all checkpoints), docs/evaluation.md (evaluation workflow, metric definitions, complete results tables), docs/datasets/README.md (DINO-WM / RT-1 / CSGO formats, downloads, splits), docs/applications/planning.md (MPC + CEM model predictive control), docs/applications/long_rollout.md (long-horizon autoregressive rollout), docs/applications/video_to_3d.md (Depth Anything 3 point cloud pipeline).
The project builds on several existing codebases, including Latte, Vid2World, DFoT, and DINO-WM, and draws design inspiration from NanoGPT, NanoChat, and Boyuan Chen's research templates. If this repository is helpful for research, you mayconsider citing it.
Limitations or information not provided: The materials do not specify the specific pricing, license details, the range of supported operating systems (only the supported Linux x86_64 environment is mentioned), minimum hardware requirements, or dataset support beyond Point Maze, DINO-WM PushT, CSGO, and RT-1; these are not yet provided.
