Researchers can study streaming rendering and long-horizon drift with measured H200 recipes; the released implementation is not yet real time. Core capabilities include Causal chunk generation, Start-image conditioning, Proxy-frame conditioning, and Six-chunk attention window.
Use the public repository and its patched Code World Model dependencies. The documented run requires an H200-class 141 GB GPU; 80 GB cards are not currently supported. Additional characteristics include Four-step distilled inference, FlashAttention-3 integration, Optional FP8 mode, and Single-H200 inference recipe.
