LingBot-World 2.0 - Ant Lingbo Technology's Open-Source Real-Time Interactive World Model

What is LingBot-World 2.0

LingBot-World 2.0 is an open-source real-time interactive world model from Ant Lingbo Technology that can be played in real time like a game. The game supports movement, attacks, spellcasting, jumping, and other actions, as well as environmental events such as rain and snow. It can generate frames continuously for hours without drift, and can reach 720P/60FPS on high-end configurations. The model has been open-sourced in four versions: 14B, 1.3B Small, Bidirectional, and Causal Pretrain, and can be used for games, embodied intelligence simulation, and interactive data generation.

Main Features of LingBot-World 2.0

Real-Time Interactive Exploration: The model generates an open world with free walking and camera rotation in real time at 720P / 60FPS with sub-second latency.
Diverse Character Actions: Supports a rich range of actions such as attacking, archery, spellcasting, shooting, jumping, and gliding.
Environment Event-Driven: Can respond to text-driven dynamic world events such as snow and rain.
Hour-Level Long-Horizon Generation: During long play sessions, the visuals remain consistent, without drift or collapse.
Agent Collaborative Control: Built-in "Player Agent + Director Agent", responsible respectively for character behavior planning and scene element synthesis.
Multiplayer Shared World: Supports multiple players simultaneously immersing themselves in the same generated world and collaboratively shaping the scene.
Embodied Intelligence Simulation: Extends from the game world to robot simulation, future state prediction, and interactive data generation.
Lightweight Local Deployment: The 1.3B Small version can run in real time on a consumer-grade single GPU.

Technical Principles of LingBot-World 2.0

Causal Pretraining Paradigm: Through carefully designed Causal Pretrain, it achieves an unlimited interactive horizon while maintaining long-term consistency in output quality, which is the core of long-horizon drift-free generation.
Real-Time Distillation: Distills a real-time variant from the base model (with the Bidirectional model as a high-quality distillation source), ensuring inference speed sufficient to drive a 720P/60FPS video stream.
Action-Conditioned Visual Dynamics: Learns "action-frame" conditional dynamics from massive first-person, synthetic, and web videos, enabling the model to faithfully ground operations and events into visual changes.
Dual-Agent Architecture: Introduces the first Agentic Harness—the Pilot Agent plans and executes character behavior, while the Director Agent synthesizes new environment elements as the scene progresses, realizing a combination of agent-driven operation and user intervention.
Large and Small Model Collaboration: The 14B main model is paired with a 1.3B lightweight model, balancing performance and single-GPU deployability while lowering the barrier to research reproduction.

Follow WeChat and reply "Open Source" to join theAI Open Source Project Exchange Group

How to Use LingBot-World 2.0

Unlimited Interactive Horizon: The causal pretraining paradigm enables models to generate continuously at an hourly scale without long-term visual drift, breaking through the duration bottleneck of traditional video models.

Real-time smooth experience: The distilled real-time variant supports 720P / 60FPS with sub-second latency, truly reaching a "playable" level of interactivity.

Rich interactive elements: Covers diverse actions such as attacking, archery, spellcasting, and shooting, as well as text-driven environmental events such as rain and snow, greatly expanding the dimensions of interaction.
Pioneering dual-agent architecture: The Pilot Agent is responsible for character behavior planning, and the Director Agent is responsible for dynamically synthesizing scene elements, with agent-driven behavior and user intervention running in parallel without conflict.
Multiplayer collaborative sharing: Supports multiple users entering the same generated world at the same time, with one person playing and one person directing, changing the traditional single-player generation mode.
Low deployment threshold: The 1.3B Small version only requires a consumer-grade single GPU to run in real time, so individual developers can also reproduce it locally.

Core advantages of LingBot-World 2.0

Ultra-long horizon: Causal pretraining supports continuous generation at the hourly scale, with no long-term visual drift.
Real-time playability: The distilled real-time variant achieves 720P/60FPS and sub-second interactive latency.
Rich interaction: Covers actions such as attacking, archery, spellcasting, and shooting, as well as text-driven environmental events.
Dual-agent architecture: The Pilot Agent plans character behavior, and the Director Agent dynamically synthesizes scenes.
Multiplayer collaboration: Supports multiple users sharing the same generated world, with one person playing and one person directing.
Low-threshold deployment: The 1.3B Small version can run locally in real time on a consumer-grade single GPU.
Diverse scenarios: Extends from game worlds to robot simulation, future prediction, and interactive data generation.

Project links for LingBot-World 2.0

Project website：https://technology.robbyant.com/lingbot-world-v2
GitHub repository：https://github.com/robbyant/lingbot-world-v2
HuggingFace model library：https://huggingface.co/robbyant/lingbot-world-v2-14b-causal-fast
arXiv technical paper：https://arxiv.org/pdf/2607.07534

Comparison of LingBot-World 2.0 with similar competitors

Dimension	LingBot-World 2.0(Ant Lingbo)	Genie 3 （Google DeepMind）
Producer	Ant Group Lingbo Technology (Robbyant)	Google DeepMind
Real-time specifications	720P / 60FPS, sub-second control latency	720p / 24fps, real-time response
Consistency duration	Hourly scalecontinuous generation without drift	About 1 minute of visual consistency
Interaction Modes	Actions such as attacking, shooting arrows, casting spells, and firing + text events + two-player collaboration (Player/Director)	Keyboard/mouse controls character movement, interaction with objects, and commands to generate objects
Multiplayer Collaboration	✅ Supports multiple players sharing the same generated world	❌ Not currently supported
Technical Approach	Causal pretraining + real-time distillation + Pilot/Director dual-Agent architecture	Autoregressive diffusion model + scalable latent space (1.5T parameters)
Degree of Openness	Fully open source: Code + four sets of 14B/1.3B weights + technical report	Closed source, available only through a Gemini Ultra subscription; Project Genie has opened invitations for game prototype testing

Application Scenarios of LingBot-World 2.0

AI Game Development: As a real-time generative game engine, it dynamically generates levels, scenes, and events based on player actions, reducing the production cost of AAA-level content.
Film and Virtual Production Previsualization: Directors can use text commands to build and roam virtual scenes in real time, quickly preview shots and atmosphere, and greatly compress the early-stage set construction and revision cycle.
Embodied Intelligence Training Ground: Provides robots with a highly realistic, interactive simulation environment to train perception, decision-making, and long-horizon interaction capabilities, reducing the cost of real-machine trial and error.
VR/AR and Digital Cultural Tourism: Generates infinitely explorable immersive virtual spaces for virtual exhibitions, cultural tourism guides, and multiplayer collaborative metaverse experiences.
Interactive Education and Training: Builds operable virtual experiments, emergency drills, and historical scenes, allowing learners to gain deeper immersive training in a "playable world" than by watching videos.