Ant Group's LingBot-World 2.0 Goes Open-Source: World Model Achieves Hour-Long Real-Time Generation with Multi-Agent Multiplayer

On July 9, Ant Group's Lingbo Technology (蚂蚁灵波) open-sourced LingBot-World 2.0 (also called LingBot-World-Infinity), a next-generation real-time interactive world model that advances world model generation from minutes to hours while maintaining 720p/60fps high-definition output. The release marks a significant leap in generation duration, dynamic interactivity, and real-time performance.

In January of this year, Lingbo open-sourced the first-generation LingBot-World model, which tackled the persistent "long-term drift" problem in extended video generation, achieving stable continuous output of nearly 10 minutes with keyboard-and-mouse control capabilities. LingBot-World 2.0 builds on this foundation with transformative improvements.

LingBot-World 2.0 adopts a causal pretraining paradigm and introduces a proprietary MoBA mechanism, allowing the model to learn how the world evolves by following real-world interaction sequences — predicting subsequent changes based on already-rendered frames, actions, and scene states. In stress tests lasting up to one hour of uninterrupted runtime, the model maintained clear textures and coherent scene structure with no significant quality degradation.

On the interaction front, LingBot-World 2.0 supports actions including attacking, archery, spell-casting, shooting, jumping, and gliding, with results generated in real-time by the model while maintaining physical plausibility and visual consistency. Users can also trigger events like day-night cycles, weather changes, and entity injections via text commands. The system features a dual-Agent architecture: a Pilot Agent that plans and executes character behavior, and a Director Agent that proposes new events in real-time as the scene progresses.

Notably, LingBot-World 2.0 supports multiple users entering and exploring the same persistent world simultaneously — an industry first for world models. To ensure real-time interactivity, Lingbo distilled a fast inference version from the pretrained model and optimized the generation pipeline, enabling stable 720p/60fps output with near-zero latency when users control characters via keyboard.

The open-source release lowers the barrier for world model adoption. Beyond applications in game content generation, film previsualization, virtual simulation, and digital twins, LingBot-World 2.0 serves as a key component of Lingbo's full-stack Brain 2.0 architecture, supporting robot and embodied intelligence training. Users can experience the model on the Reactor platform or the LingGuang app.