H3-World Turns MiniMax-H3 into World Model with Only 8K Samples
linoy_tsaban · x · 2026-09-02
H3-World transforms the MiniMax-H3 pretrained language model into a world model without adding new action modules. By converting keyboard controls into textual instructions and injecting them via the model's text pathway, it achieves language-native and temporally grounded control. The method is highly efficient, using only 8,000 gameplay samples, 10,000 LoRA steps, and 0.199% trainable parameters to control character and camera motion, even on unseen action compositions.
More from Research
- Open source approach aligns with World Labs' Atlas release — mickmumpitz · 2026-09-02
- View: Different tasks rely on sampling bottleneck, not just depth limits — voooooogel · 2026-09-02
- Prediction: Mechanistic Interpretability Will Surpass CoT Monitoring — tszzl · 2026-09-02
- Why looping middle layers isn't enough: computational graph depth matters — _AndrewZhao · 2026-09-02
- JIT-Agent Paper: Dynamic Harnesses Let Small Models Beat Giants While Cutting Costs by 36% — rohanpaul_ai · 2026-09-02
- LoRA Co-inventor Joins Mercor, Releases 397B RL Training Guide — himanshustwts · 2026-09-02