Open Dreamer rebuilds Dreamer 4 in JAX with full training recipe
Research group Reactor has released a JAX/Flax reproduction of DeepMind's Dreamer 4 pipeline, publishing the full recipe and the stability fixes that made it train.
A research group called Reactor has released Open Dreamer, an open-source JAX/Flax reproduction of the Dreamer 4 world-model pipeline — the DeepMind system that learned to mine diamonds in Minecraft by training an agent entirely inside its own learned model of the game. Where the original paper shipped without a full public implementation, Open Dreamer publishes the complete training stack: a causal video tokenizer, an action-conditioned latent dynamics model, the full training recipe with configurations, an inference harness for local rollouts, and a browser-based Minecraft demo generating frames in real time.
What's in the release — and what isn't
The reproduction centers on a 1.6-billion-parameter dynamics transformer (30 layers, model width 1,920) trained for 200,000 steps with the Muon optimizer, running 256 frames per GPU with roughly 24GB of model state and sustaining a reported 57–58% model-FLOPs utilization on NVIDIA B200 hardware. Notably absent are pretrained weights and the behavior-cloning and reinforcement-learning stages — this is the world-model backbone, not a turnkey agent.
Stability, not throughput
The most valuable part of the release may be its engineering notes. The team reports that training stability, not speed, was the real bottleneck, and documents six critical fixes for a failure mode that rarely makes it into papers: configurations where the loss curve kept improving while actual generation quality quietly degraded. That decoupling of loss from output quality is a known trap in world-model training, and a written recipe for avoiding it is scarce.
The release lands in the middle of the industry's world-model wave — Genie, Cosmos, Marble, and their peers — where nearly all frontier work remains closed or partially released. Independent reproductions with honest failure documentation do for world models what open reimplementations once did for LLM training: they turn a handful of labs' tacit knowledge into public infrastructure. For researchers without DeepMind-scale resources, a verified recipe that trains stably is worth more than another benchmark chart.