There was no laboratory behind Diego Marti Monso and his two collaborators when they decided to reconstruct Google DeepMind’s frontier world model, Dreamer 4. Although Google’s research lab published a technical report about the model, the company never released the code or the weights. Regardless, three researchers undertook the bold tasks of reproducing a frontier model with a mere fraction of the resources.
They succeeded. Open Dreamer is an independent project co-led by Diego Marti Monso, who is best known for his work on Diffusion Forcing, as well as by Francesco Sacco and Edward Hu. The project was sponsored by Reactor.
The starting point was Dreamer 4, a world model developed by Google DeepMind researchers Danijar Hafner, Wilson Yan, and Timothy Lillicrap. Introduced in September 2025, Dreamer 4 is a scalable AI agent and world model that learns to solve complex control tasks entirely through novel internal "imagination training." A crowning achievement, Dreamer 4 learned to obtain diamonds in Minecraft without ever interacting with the live game during training. This task requires perception, exploration, long-term planning, and interaction—things that existing AI models have historically struggled with. Dreamer 4 received coverage in publications such as the TalkRL podcast and Tech Xplore.
Marti Monso and his collaborators developed the model and training code from scratch. Together, they scaled the system to a 1.6 billion-parameter model using high-performance computing clusters and thoughtful algorithm building. The three researchers rebuilt critical frontier technology that was previously inaccessible to the research community.
Reproducing a frontier result from a paper alone is different from implementing a documented method. "There was no answer key," said Marti Monso. "When something failed, we had to figure out whether the problem was in the code, the training setup, or our understanding of the method itself." The scale of the project made mistakes costly in terms of time and computing resources. A small error might only become apparent after a lengthy training run had already consumed substantial resources. The three researchers were working without an institutional lab or its resources. Much of the project consisted of determining which choices were essential, which could be inferred from the paper, and which had to be settled empirically. This reconstruction work would otherwise have had to be repeated independently by every subsequent team.
Marti Monso has a specific view of what reproducible research should entail. While publishing the weights of a trained model can show researchers what the system is capable of doing, it does not explain the training process or the most important factors and decisions when building the model. Marti Monso argues that the full recipe is important because it provides other researchers with the necessary information to reproduce the process and analyze the choices that produced the result.
"If you only know the final result, you still have to reconstruct the entire path yourself," Marti Monso said. "Publishing the recipe gives another researcher something they can actually reproduce." This principle became central to Open Dreamer. Rather than limiting the project to a description of the finished system, the team released the model and training code.
Monso described this approach as surpassing an open-weights release because the training process itself was made available. The trained world model can also be run interactively in a web browser, providing researchers and other users with a direct way to examine its behavior. This kind of open-source work is crucial, particularly in a nascent field such as world modeling, where researchers are still developing their intuition.
The project also revealed the practical challenges of reproducing cutting-edge research without the support of an established laboratory. Training large models requires substantial computing resources, which are unavailable to anyone outside major AI labs. Marti Monso and his collaborators had to solve these problems with limited resources, unlike the original Dreamer 4 authors.
Additionally, the lack of implementation details made investigating each unresolved problem more expensive. The team had to determine which choices were essential, which could be reconstructed from the published description, and which assumptions needed to be tested through training. This experience reinforced their decision to document their process, rather than leaving future researchers with the same missing information.
Open Dreamer received significant attention after its release. As of August 2026, the project had received over 350 GitHub stars, becoming a trending topic on X with over 1 million views. It was also featured by MarkTechPost. This is not surprising, given the current interest in world models. Many tech leaders see world models as crucial to the future of robotics and video generation. More importantly, the field of world models gained another significant resource.
With Open Dreamer's publication, any researcher can build on the Dreamer 4 algorithm (or its equivalent). This model can be extrapolated to robotics, Minecraft and more and can now be used by the broader research community to pursue loftier goals. Marti Monso, Sacco, and Hu deserve the credit.