Key Info

Tencent open-sourced two repositories addressing steering and physics in world models: Hy-Nomas and Looking-Ahead.

Highlights

  • Hy-Nomas accepts multi-turn text and camera input, generating the world segment by segment with a KV cache that retains history.
  • It runs on a single RTX 4090 using an INT8 low-memory configuration, and Diffusers support has landed.
  • The community has already integrated Hy-Nomas into ComfyUI, SGLang, and WanGP.
  • Looking-Ahead targets autoregressive models that rely on local continuity: during training the model predicts chunks further ahead, and those branches are dropped at inference.
  • It uses the same 5B base model and same data.