Ant Group's CodeWorld Model 2.0 streams interactive worlds for 30 minutes without drift

机器之心 · wechat · 2026-11-18

Ant Group, Westlake University and Zhejiang University released CodeWorld Model 2.0, whose Invoker-Render engine separates world-state management (code + structured proxies) from visual generation. It supports five proxy types from sparse 3D boxes to skeletons, and the ProForcing three-stage training enables 30-minute continuous streaming generation without appearance drift. Its single-node 8-GPU SGLang engine cuts forward calls per block from 6 to 4. In 140 blind pairwise evals, it ranks first on all four Elo metrics (overall 1343.2), beating the strongest baseline with 62.9% win rate and +97.9 Elo. Models, inference engine and tools are open-sourced.

Related event: Ant's Inclusion AI Releases Code World Model 2.0 with Invoker-Render(2 posts)→

Original post →

More from Multimodal

Multimodal channel →