Researchers from Hugging Face and affiliated labs have unveiled ABot-World-0, an action-conditioned video world model designed for real-time, long-horizon closed-loop interaction. The model can run on a single desktop GPU, achieving up to 16 FPS at 720P resolution on an NVIDIA RTX 5090, with an action-to-first-frame latency of just 1.2 seconds and peak VRAM usage of approximately 19 GiB.
The system is built on a multi-source data infrastructure that includes AAA games, simulation engines, and internet videos to learn controllable world dynamics. A key component is WorldExplorer, which performs agent-driven data collection guided by training feedback. The pipeline applies 14 deterministic quality checks, VLM-based assessment, and synchronized action and text annotation.
ABot-World-0 uses a progressive distillation approach: a bidirectional action-conditioned teacher is distilled into a causal student via teacher forcing and ODE distillation. The novel LongForcing technique aligns long student self-rollouts with an extended-horizon teacher, mitigating accumulated distribution shift and autoregressive drift. Raw keyboard actions serve as a unified control interface for scene roaming and third-person character interaction, while reference-character memory ensures identity consistency during rollouts.
For deployment, the team co-designed a streaming inference stack featuring a lightweight VAE decoder, efficient attention mechanisms, memory-aware scheduling, and low-bit DiT inference. Experiments on WorldRoamBench and extended interactive rollouts demonstrate competitive controllability and coherent long-horizon world evolution. The project is open-source, with code, model, and interactive demo available on GitHub and Hugging Face.