Jasmine: A Simple, Performant and Scalable Jax-Based World Modeling Codebase (arxiv.org)

Jasmine is a new open-source JAX-based world modeling codebase built to make training scalable, reproducible, and fast. The authors show it scales from a single host to hundreds of accelerators with minimal code changes, and—thanks to focused performance optimizations across data loading, training loops, and checkpointing—reproduces the CoinRun case study about an order of magnitude faster than prior public implementations. Jasmine also enforces fully reproducible training runs and supports a range of sharding configurations, making it practical to move experiments from laptop-scale debugging to large multi-accelerator runs without rewriting core logic. For the AI/ML community, Jasmine fills a gap in open training infrastructure for world models, which are increasingly important for sample-efficient learning in robotics and other data-scarce domains. Technically, the codebase leverages JAX’s parallelism primitives and careful I/O/serialization design to reduce bottlenecks that typically block scale-up; it pairs with curated large-scale datasets to enable rigorous benchmarking across model families and architectural ablations. The result is an accessible platform for researchers to iterate quickly, compare architectures reproducibly, and push world-model research at production-like scale.