Alibaba Releases ABot-World-0, a 5B Interactive World Model Running at 720P Real Time on One GPU
AI video model releases
Alibaba Releases ABot-World-0, a 5B Interactive World Model Running at 720P Real Time on One GPU
Alibaba's AMAP Computer Vision Lab released ABot-World-0, a 5 billion parameter action conditioned world model generating explorable 720P environments at 16 frames per second with 1.2 second latency on a single RTX 5090 GPU using 19GB VRAM.
The model uses a training technique called LongForcing that targets long horizon stability as an explicit training objective, reducing the scene repetition and quality collapse that typically break extended navigation sessions.
ABot-World-0 is built on Alibaba's Wan2.2-TI2V backbone, generating one frame per user action using the prior frame as a visual anchor rather than producing a full video clip at once.
Weights are released in SafeTensors format on HuggingFace as acvlab/ABot-World-0-5B-LF under an Apache 2.0 license, permitting commercial use without extra licensing review.
A companion model, ABot-3DWorld-0, uses 3D Gaussian Splatting to turn prompts into persistent spatial environments instead of video sequences, and both are covered in an arXiv paper submitted July 25, 2026.
cho.sh