Ant Group’s AI Unit Open-Sources Hour-Long Interactive World Model Ant Group's dedicated embodied-intelligence unit, Robbyant , has offi...
Ant Group’s AI Unit Open-Sources Hour-Long Interactive World Model
Ant Group's dedicated embodied-intelligence unit, Robbyant, has officially released LingBot-World 2.0. This major launch introduces an advanced interactive world model capable of generating highly coherent video environments for up to a full hour. Operating at a crisp 720p resolution and a smooth 60 frames per second (fps), the system establishes a new benchmark for duration and performance in simulated digital environments.
By making this technology widely available, the organization aims to democratize access to high-fidelity environmental simulation, providing developers and AI researchers with the core infrastructure needed to train and test advanced systems within highly stable, long-horizon virtual spaces.
Hour-Long Interactive Worlds: Inside the 'Infinity' Upgrade
The debut of LingBot-World 2.0 represents a substantial technical leap from the capabilities of the original LingBot-World architecture, which was open-sourced earlier this year in January. While the initial iteration of the framework supported video generation sequences capped at roughly 10 minutes, the newly unveiled version expands this operational window sixfold.
Subtitled "Infinity," this upgraded architecture introduces fundamental enhancements to both world prediction capabilities and real-time user interactivity. Rather than simply rendering passive video playbacks, the system allows users to actively navigate and execute precise actions inside the generated environments in real time.
The underlying engine continuously calculates environmental responses based on user inputs, maintaining visual logic and structural consistency across the entire 60-minute duration. This prevents the spatial drift, visual degradation, and loss of environmental context that typically plague long-form generative video pipelines.
To maximize accessibility and foster collaboration across the global AI ecosystem, the entire model repository has been made publicly available on GitHub and Hugging Face under an open-source license. Recognizing the computational demands of hosting such long-horizon models, the release features day-zero support for SGLang, a specialized serving framework engineered specifically for executing large generative models with low latency and high throughput. This integration ensures that engineering teams can deploy the model into production environments immediately without needing to build custom serving pipelines.
LingBot-Video: A MoE Foundation Model Built for Embodied Intelligence
Alongside the core interactive world model, Robbyant has simultaneously open-sourced LingBot-Video. The development team describes this release as the first open-source video generation foundation model built entirely on a Mixture-of-Experts (MoE) architecture designed explicitly for the demands of embodied intelligence.
Standard dense architectures often struggle to balance the intense computational requirements of high-frame-rate video synthesis with the real-time processing speeds required by physical robotic systems. The MoE framework addresses this constraint by dynamically routing information only to specialized subsets of the neural network during inference.
[Input Token] ---> [Router] ---> [Expert 1 (Active)] ---> [Output]
---> [Expert 2 (Inactive)]
The model features a 13-billion-parameter total capacity, yet it strategically activates only 1.4 billion active parameters per token. This architectural design yields significant improvements in inference efficiency, reducing the overall computational footprint while simultaneously enhancing the physical plausibility of the generated scenes.
For robotics applications, physical plausibility is crucial. The model correctly interprets and simulates complex physical properties, environmental dynamics, and spatial interactions, ensuring that the visual data produced closely mirrors real-world physics.
To lower the barrier to entry for developers and researchers, the comprehensive release includes the full model weights, the underlying codebase, and seamless integration out of the box with popular ecosystem tools, including the Diffusers library and the SGLang serving framework.
A Growing Open-Source Stack for Advanced Robotics
These twin releases cap a highly prolific stretch of development for Robbyant. Throughout 2026, the unit has consistently open-sourced a series of critical embodied-AI components designed to function as an interconnected ecosystem.
This rapid release schedule includes the launch of LingBot-Map in April, which introduced real-time streaming 3D reconstruction capabilities, followed by the deployment of LingBot-Depth 2.0 earlier this month, a system focused on enhancing spatial perception and depth estimation for autonomous systems.
Industry analysts, including commentators at The Next Web, have noted that the sheer velocity of these sequential rollouts effectively amounts to "a full embodied-AI stack" released in the span of a single week. By providing the tools for spatial mapping, depth perception, foundational video generation, and interactive world modeling, the unit offers an end-to-end framework for modern robotics research.
Developers looking to explore the capabilities of these new models can access the technology through multiple channels. An online interactive demo is currently available on the Reactor platform for immediate testing, while the complete, unmodified repository can be cloned directly from GitHub for local deployment, customization, and fine-tuning.