OpenDreamer

OpenDreamer is an open-source world model and training codebase that reproduces the Dreamer 4 paper, targeting a playable Minecraft world model. It was written by Diego Marti Monso, Francesco Sacco and Edward Hu, with support from Reactor, and published with a blog post titled "How to train a frontier-level world model" dated July 21, 2026.

The architecture has two parts: a transformer-based masked autoencoder tokenizer (about 100x compression, no KL or adversarial loss) and a dynamics model built on a block-causal transformer with space layers and causal time layers. The dynamics model is trained with diffusion forcing, flow matching and shortcut models, and also predicts actions so that the world model and policy share the same representation. The authors prototyped the pipeline on CoinRun on a single GPU before scaling to Minecraft.

The release includes the model, training code and the lessons and failure modes from scaling. The released repository does not include the behaviour-cloning or RL code, since it was not used for Minecraft. A live playable demo runs on the Reactor runtime. Vendor/organisation details beyond the "next-state" GitHub namespace are not stated in the provided material.