Robbyant releases interactive model on Reactor, GitHub and Hugging Face

Robbyant, an embodied AI company within Ant Group, has released LingBot-World 2.0 (Infinity) as an open-source world model.

From Shanghai, the company announced a system that supports hour-long continuous generation and 720p/60fps real-time output.

Users can access LingBot-World 2.0 on Reactor, where content generates, transmits and displays at the same time.

Because of that pipeline, users do not need to wait for a full sequence to finish before interaction starts.

Version 1.0 delivered minutes-level stable generation, while LingBot-World 2.0 extends that to hour-long output.

According to Robbyant, the upgrade improves both world prediction and interactivity, and adds a native agent mechanism.

Robbyant described that mechanism as moving generated worlds from watchable and controllable to “sustainably interactive and dynamically evolving”.

Engineers built the model with a Causal Pretraining Paradigm and a proprietary MoBA, or Mask of Bidirectional Attention, mechanism.

That approach teaches the model to learn world evolution in chronological order during generation.

As a result, LingBot-World 2.0 aims to prevent compounding errors such as texture blurring, geometry collapse and scene breakdown.

In hour-long stress tests, Robbyant reported zero quality drift while maintaining visual fidelity.

Alongside the pre-trained release, the company also distilled a fast inference version and optimised the generation pipeline.

Reactor real-time controls

On Reactor, LingBot-World 2.0 delivers 720p/60fps visuals with low-latency feedback and keyboard controls.

Users can move a character or switch perspectives in real time through those controls.

Beyond navigation, the action set includes attacking, shooting arrows, casting spells, jumping and gliding.

Outcomes change with the live scene state, which keeps actions physically plausible and visually consistent.

Text commands can also trigger day-night cycles, weather changes and entity injection inside the generated world.

Inside the system, a Pilot Agent plans and executes character behaviour while a Director Agent introduces new events.

Meanwhile, LingBot-World 2.0 supports multiple users in a single persistent world for collaborative exploration.

Robbyant said that multiplayer persistence marks a critical step toward AI-native multiplayer experiences.

The release is available through Reactor, GitHub and Hugging Face.