H3-World Turns MiniMax-H3 into World Model

Xingyi Yang, Assistant Professor at PolyU, posted that the team shares H3-World. The post states this is the first effort to turn MiniMax-H3 itself into a world model. It claims no new action module was added and that H3’s pretrained language understanding was converted directly into world control. The same post notes the method used 8K samples and 0.199% trainable parameters. A 49-second video attachment opens with split-screen fantasy scenes including snowy mountains, desert oasis, neon cyberpunk city, and jungle temple. Researcher Ricky T. Q. Chen retweeted the announcement.