LingBot-World 2.0

Model Overview

LingBot-World 2.0 (also known as LingBot-World-Infinity) is a groundbreaking, open-weights interactive world model released on July 9, 2026. Developed by Robbyant, the embodied AI research lab within Ant Group, this model pushes the boundaries of real-time video generation. Unlike previous models limited to seconds or minutes of output, LingBot-World 2.0 allows for continuous, hour-long real-time interactive generation without scene collapse. Released under the Apache 2.0 license, it is designed to be self-hostable and accessible to the broader research community.

Capabilities

LingBot-World 2.0 introduces several revolutionary features for interactive AI:

• Infinite Interaction Horizon: Capable of generating stable, continuous 720p interactive video at 60fps for over an hour with zero quality drift or geometric breakdown, utilizing a novel Mask of Bidirectional Attention (MoBA).
• Dual-Agent Architecture: Features a native dual-agent mechanism:
  • Pilot Agent: Manages and executes specific character behaviors and player inputs.
  • Director Agent: Dynamically introduces new environmental events and changes to keep the scene evolving.
• Real-Time Interactivity: Users can navigate the generated world using standard WASD keyboard controls and perform complex actions (e.g., attacking, shooting, spell-casting) starting from just a single input image.
• Multiplayer Persistence: Supports persistent environmental states, allowing for multiplayer-like experiences within a single generated instance.

Example Use Cases

• Next-Generation Gaming: Serving as the foundation for infinite, dynamically generated video games where environments and narratives evolve in real-time based on player actions.
• Embodied AI Training: Providing an infinitely scalable, highly realistic, and physically consistent interactive simulator for training robotic systems and autonomous agents.
• Interactive Storytelling: Enabling creators to build "choose-your-own-adventure" cinematic experiences where viewers can take control of the narrative in real-time.

Performance & Benchmarks

While standard benchmark scores were not highlighted in its initial metadata, LingBot-World 2.0 represents a massive performance leap in efficiency. Robbyant distilled a faster inference variant from the base 14B parameter model, enabling this complex, 60fps real-time interactive generation to run entirely on a single consumer-grade GPU.

Intended Use & Limitations

LingBot-World 2.0 is intended for developers, game designers, and AI researchers exploring embodied AI and interactive world synthesis. Because it is an open-weights model, it can be self-hosted and fine-tuned. Limitations may include high VRAM requirements for the full-sized model, and the inherent unpredictability of the Director Agent in highly constrained or specific narrative use cases.

About Robbyant

Robbyant is the dedicated embodied AI research laboratory of Ant Group. They focus on bridging the gap between digital generative models and physical robotics. LingBot-World 2.0 is part of a broader embodied-AI stack developed by the lab, which includes vision-language-action models like LingBot-VLA 2.0 and LingBot-Vision, all aimed at advancing the future of interactive and physical artificial intelligence.