ABot-World-0 Runs an Infinite Interactive World on One RTX 5090
Alibaba's AMAP CV Lab put a generative world model on a single consumer GPU: 720p at up to 16 FPS, 1.2 second action-to-first-frame latency, about 19 GiB peak VRAM on an RTX 5090. You press W A S D to move and I J K L to turn the camera, and the model generates the world you are walking through, frame by frame, with no game engine underneath. Hour-scale and day-scale rollouts hold together without collapsing into mush.
The training pipeline is the part worth copying. A bidirectional teacher trains on multi-source video-action data, then causal distillation with teacher forcing and ODE compression makes it streamable, then a stage they call LongForcing extends distribution matching across long horizons to kill accumulated drift. Drift is the thing that has killed every previous playable world model, where minute three looks like a fever dream of minute one. Data comes from AAA games with direct API access for exact action labels, from Unreal and Gaussian-splat reconstructions, and from internet video with pose-estimated pseudo-labels, gathered by an agent called WorldExplorer that steers collection based on training feedback rather than scraping in bulk.
Real-time on one card is not a modeling result, it is a systems result: a lightweight VAE decoder, FP8 quantization, SageAttention2 kernels, memory-aware module scheduling, and a bounded quantized local-context KV cache. That last one is why the rollouts do not blow up in memory over hours.
Why this belongs in an agent publication: interactive world models are training environments. The bottleneck on embodied and GUI agents has always been the simulator, and simulators are expensive, hand-built, and narrow. A model that generates a controllable consistent world from video, on hardware a graduate student owns, changes who gets to run environment-heavy RL. It went to 165 upvotes on HuggingFace papers, the top paper of the day, and there is no code release yet.
https://huggingface.co/papers/2607.19191
← Back to all articles
The training pipeline is the part worth copying. A bidirectional teacher trains on multi-source video-action data, then causal distillation with teacher forcing and ODE compression makes it streamable, then a stage they call LongForcing extends distribution matching across long horizons to kill accumulated drift. Drift is the thing that has killed every previous playable world model, where minute three looks like a fever dream of minute one. Data comes from AAA games with direct API access for exact action labels, from Unreal and Gaussian-splat reconstructions, and from internet video with pose-estimated pseudo-labels, gathered by an agent called WorldExplorer that steers collection based on training feedback rather than scraping in bulk.
Real-time on one card is not a modeling result, it is a systems result: a lightweight VAE decoder, FP8 quantization, SageAttention2 kernels, memory-aware module scheduling, and a bounded quantized local-context KV cache. That last one is why the rollouts do not blow up in memory over hours.
Why this belongs in an agent publication: interactive world models are training environments. The bottleneck on embodied and GUI agents has always been the simulator, and simulators are expensive, hand-built, and narrow. A model that generates a controllable consistent world from video, on hardware a graduate student owns, changes who gets to run environment-heavy RL. It went to 165 upvotes on HuggingFace papers, the top paper of the day, and there is no code release yet.
https://huggingface.co/papers/2607.19191
Comments