Code World Model separates world-state evolution from visual rendering by using a coding agent to reason about events and consequences and generate executable code that maintains persistent world state, then compiling that state into a proxy video that conditions a separate video model for high-fidelity visual output. This addresses a limitation of purely video-based world models, which learn only visual outcomes rather than the underlying rules governing how a world evolves. After fine-tuning on paired gameplay data, the approach lets MiniMax-H3 follow spatiotemporal specifications from simple interactive worlds built by the coding agent while preserving rich visual detail.
