Summary

LingBot-World 2.0 (LingBot-World-Infinity) is an advanced open-source project addressing the creation of dynamic, interactive virtual worlds. It targets researchers and developers seeking to build complex simulated environments with consistent, high-quality outputs and real-time responsiveness. The project offers an unbounded interaction horizon, rapid response times suitable for high-framerate video streams, and a novel agentic harness for sophisticated character and environmental behaviors, significantly enhancing the richness and interactivity of simulated worlds.

How It Works

The system employs a causal pretraining paradigm to achieve an unbounded interaction horizon, ensuring consistent output quality over extended interactions. A distilled real-time variant guarantees rapid response times, capable of driving 720p video streams at 60 fps. A key innovation is the agentic harness, which integrates a pilot agent for planning and executing character behaviors and a director agent responsible for synthesizing novel environmental elements dynamically. This architecture enables highly diverse interactive elements, including a broad spectrum of actions and richer text-driven events.

Quick Start & Requirements

This codebase is built upon Wan2.2.

• Installation: Clone the repository (git clone https://github.com/robbyant/lingbot-world-v2.git), navigate into the directory, and install dependencies (pip install -r requirements.txt). Requires torch >= 2.4.0. Install flash-attn separately (pip install flash-attn --no-build-isolation).
• Models: The lingbot-world-v2-14b-causal-fast model (14B parameters) is available via HuggingFace (https://huggingface.co/robbyant/lingbot-world-v2-14b-causal-fast) and ModelScope (https://modelscope.cn/models/Robbyant/lingbot-world-v2-14b-causal-fast).
• Inference: Use generate.py for causal inference with KV caching. Multi-GPU is recommended for causal_fast inference (e.g., torchrun --nproc_per_node=8 generate.py ...). A run_fast.sh script is also provided.
• Demo: Interactive demos are available on Reactor (Web) and LingGuang (Mobile). More demos can be found on the Project Page (https://technology.robbyant.com/lingbot-world-v2).
• Tech Report: Available at https://arxiv.org/abs/2607.07534.

Highlighted Details

• Achieves an unbounded interaction horizon with consistent output quality.
• Real-time variant capable of driving 720p video streams at 60 fps.
• Introduces a diverse range of interactive elements and text-driven events.
• Pioneers an agentic harness for world modeling, featuring pilot and director agents.

Maintenance & Community

The project is developed by the Robbyant Team. Support is acknowledged from Reactor and LingGuang. Future releases are planned for additional model variants (causal-pretrained, bidirectional, 1.3B models). No direct community links (e.g., Discord, Slack) are provided in the README.

Licensing & Compatibility

The project is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0). This license strictly prohibits commercial use. Derivative works must be distributed under the same license.

Limitations & Caveats

Deployment code is explicitly not planned for release; users must refer to external projects like SGLang or flashdreams for self-deployment. Several model variants, including the causal-pretrained 14B and all 1.3B models, are marked as "TODO" and are not yet available. The CC BY-NC-SA 4.0 license restricts usage to non-commercial applications.