Alibaba's Amap Unveils ABot-World AI Models That Generate Interactive Video and 3D Scenes for Hours on a Single GPU

BEIJING, July 16, 2026 /PRNewswire/ -- Amap, Alibaba's location-based services platform, has released a pair of generative AI models that can produce interactive video and three-dimensional scenes in a unified framework, with the ability to run continuously for more than an hour on a single consumer-grade graphics processor.

The models, including ABot-World-0 and ABot-3DWorld-0, are designed to overcome key limitations in the emerging field of "world models"-AI systems that simulate physical environments. Existing video-generation models typically degrade after about one minute of inference, while 3D-scene generators produce high-fidelity assets but cannot link multiple scenes into a continuous explorable space

To address the video stability issue, the ABot-World model integrates scene navigation and character control into a single training objective, improving response accuracy to user inputs. It also applies a temporal-consistency algorithm that mitigates error accumulation during auto-regressive inference, allowing stable generation for over an hour without quality drop. The architecture imposes no built-in upper limit on inference duration.

For 3D generation, the ABot-3DWorld-0 model shares the same inference pathway as Amap's previously released ABot-Earth0.5, enabling a single architecture to produce 3D Gaussian Splatting (3DGS) assets across indoor, street-level, and aerial city scales. Outputs feature photorealistic textures and perceptible geometric structures with physical boundaries. A modular "generate-evaluate-repair" pipeline ensures quality control over long-duration scene synthesis.

Amap integrates the video-generation and 3D-scene generation models into a single product, allowing video outputs and 3DGS assets to be converted and stored jointly. Amap also introduces a "spatial teleport" mechanism that lets users create anchor points between any two generated 3D worlds. When traversed, the system seamlessly unloads the current scene and loads the target while maintaining smooth transitions in perspective and lighting – effectively weaving isolated scenes into an expandable exploration network. Each scene can be saved, shared, or reused independently, forming an open public asset library.

On the engineering side, Amap said all inference tasks, both video and 3DGS generation, can run locally on a single consumer-grade GPU without requiring distributed computing or cloud accelerators. Model weights, inference code and deployment guides have been partially open-sourced, with a test interface available via the ABot-World Studio. Users can input text or an image to generate an AI world that supports real-time interaction, persistent storage and cross-scene teleportation.

Potential applications include simulation environments for embodied AI, rapid storyboard expansion for film pre-visualization, and immersive first-person experiences for cultural tourism and education – shifting users from passive viewers to active explorers, the company said.

The test portal is accessible at https://abot-world.amap.com/#world, with GitHub and Hugging Face mirrors also available.

SOURCE Amap