Ant Group's LingBot Open-Sources LingBot-World 2.0: Infinite Real-Time Interactive World Model with Hour-Long Stability

Ant Group's AI research division LingBot has officially open-sourced LingBot-World 2.0 (also branded as LingBot-World-Infinity), a next-generation real-time interactive world model designed to serve as a sustainable open foundation for game development, interactive content creation, autonomous driving simulation, and embodied intelligence training.

Expanded Action Spaces and Agent-Driven Worlds

Unlike conventional interactive world models limited to camera movements and basic navigation, LingBot-World 2.0 dramatically expands the action space. The model supports complex combat maneuvers—including spellcasting, melee attacks, archery, and shooting—alongside athletic movements like jumping and gliding. These are not pre-rendered animations but dynamically generated visual changes that maintain physical consistency with the environment.

The system introduces sophisticated text-driven event control, allowing users to trigger macro-environmental shifts (transitioning from cities to forests, summoning blizzards, spawning flocks of birds) or inject new entities on the fly. Two built-in AI agents manage world evolution: the Pilot Agent handles character behavior planning and execution, while the Director Agent proposes real-time environmental events, ensuring the world evolves dynamically rather than remaining static.

Critically, LingBot-World 2.0 supports multiplayer co-existence, enabling multiple users to explore and interact within the same persistent world simultaneously—marking a significant step toward AI-native multiplayer experiences.

Breaking the Time Barrier: Hour-Long Stable Generation

A persistent challenge in interactive world models is "long-term drift"—the accumulation of errors over time that leads to texture blurring, geometric distortion, and eventual scene collapse. Traditional models typically stabilize for only seconds to minutes before degrading.

LingBot-World 2.0 addresses this through a Causal Pretraining Paradigm that fundamentally suppresses compound error generation. The architecture incorporates MoBA (Mixed Bidirectional and Auto-regressive attention mask mechanism), which mitigates error accumulation and quality degradation inherent in long-range autoregressive generation while maintaining output fidelity.

In stress tests exceeding one hour of continuous operation, the model maintained sharp textures and coherent scenes without significant quality decay. The developers assert this is a structural capability rather than a short-clip artifact, positioning LingBot-World 2.0 as the only general-purpose interactive world model currently capable of hour-level—or theoretically infinite—generation.

Real-Time Performance at 720p/60fps

Achieving high-fidelity rendering with real-time responsiveness requires solving the computational tension between video quality and interaction latency. LingBot-World 2.0 employs Consistency Distillation and Distribution Matching Distillation (DMD) to create a fast, interaction-optimized variant from the base causal world model, drastically reducing sampling costs while preventing visual degradation from long-term autoregression.

Engineering optimizations include compiler-level attention kernel optimization, hybrid parallel inference strategies, dynamic KV cache scheduling, and asynchronous streaming transmission. Rather than waiting for complete video chunks, the system generates, decodes, and transmits simultaneously.

The result is stable 720p resolution at 60 frames per second, allowing users to control character movement and camera perspective via keyboard with minimal latency.

Open Source and Ecosystem Availability

The released 14B parameter model is available under a non-commercial open-source license, providing high-quality generation suitable for research validation and complex interactive world construction. The complete model weights and inference code have been made publicly accessible.

For immediate access without local deployment, Ant Group has partnered with Reactor (PC client) and Lingguang APP (mobile) to offer hosted experiences. Additionally, SGLang has provided Day-0 support for LingBot-World 2.0, enabling developers to deploy efficient inference services out of the box.

By open-sourcing this infrastructure, Ant Group aims to establish a robust foundation for next-generation interactive AI applications, from open-world gaming simulations to robotics training environments.