H3-World: Turning MiniMax-H3 Into a World Model With Just 8K Samples

MiniMax (official) ·

Key Info

H3-World, shared via MiniMax, claims to be the first project to turn MiniMax-H3 itself into a world model. It adds no new action module, instead converting H3's pretrained language understanding into world control, using only 8K samples and 0.199% trainable parameters.

Highlights

  • No separate action module: the approach directly repurposes MiniMax-H3's existing language understanding for world control.
  • Extremely lightweight: only 8K training samples and 0.199% trainable parameters are needed.
  • The paper and code are publicly available for further exploration.
Loading...