H3-World turns MiniMax-H3 33B video model into a world model

H3-World, a new preprint from Danze Chen and colleagues, converts the 33B-parameter MiniMax-H3 video generator into an interactive world model by training just 0.199% of the parameters. The paper on arXiv, submitted 1 September 2026, reports that 8,000 gameplay samples and 10,000 LoRA optimization steps were enough. No dedicated action modules were bolted onto the base model.

The setup encodes actions as "structured combinations of character and camera instructions aligned with video latents," then uses what the authors call "temporal attention routing, which restricts each instruction to its intended time interval" to keep one action from bleeding into the next. The underlying claim is that a large video generator already carries the semantic scaffolding for world control, so a thin adapter is enough to surface it.

The authors write that the framework achieves "precise, temporally grounded world control, without introducing dedicated action modules," and that it generalizes to unseen scenarios. The abstract publishes no head-to-head numbers against existing world-model baselines, no wall-clock inference cost on the 33B backbone, and no external replication. This is a single preprint from a small team.