Reproducing the Dreamer 4 Pipeline
A research team named Reactor has released Open Dreamer, an open implementation of the world-model pipeline written in JAX and Flax NNX. The project includes two main repositories: one holding the training pipeline with a causal video tokenizer and action-conditioned latent dynamics model, and another providing a minimal local rollout harness. A browser demo streams a generated Minecraft world in real time, showcasing the model's capabilities.
Architecture and Training Stability
The dynamics model contains 1.6 billion parameters and uses a block-causal transformer backbone. The tokenizer is a transformer-based Masked Autoencoder, achieving roughly 100x compression without requiring KL or adversarial loss. The team reported 57 to 58 percent model FLOPs utilization on B200 hardware, attributing stability improvements to the Muon optimizer, strict mixed precision boundaries, and mandatory EMA weights for diffusion inference.
Open Source Foundations for Embodied AI
The stated objective was to reproduce the Dreamer 4 pipeline while deliberately avoiding methods outside the original research paper to keep the search space narrow. While the repository does not yet include the full behavior-cloning or reinforcement learning training loop, it provides a stable, reproducible foundation for the embodied AI research community to build upon.