LingBot-World is an open-source world simulator derived from video generation technology, positioned as a top-tier world model. The project has released LingBot-World-Infinity, the original repository is no longer actively maintained, and subsequent updates have moved to https://github.com/Robbyant/lingbot-world-v2。
Purpose and Positioning As an open-source world model, LingBot-World aims to narrow the gap between open-source and closed-source technologies, empowering the community in practical applications such as content creation, gaming, and robot learning.
Main Capabilities
- High-fidelity and diverse environments: Maintains high fidelity and robust dynamics across a wide range of environments, including photorealistic, scientific scenes, and cartoon styles.
- Long-term memory and consistency: Supports minute-level time spans while maintaining contextual consistency (i.e., long-term memory).
- Real-time interaction and open access: Supports real-time interaction, generating at 16 frames per second with latency below 1 second. Provides public access to code and models.
- Model versions: LingBot-World-Base (Cam), LingBot-World-Base (Act), LingBot-World-Fast (supports KV cache acceleration for causal inference, processing video frames in chunks).
- Quantized version: The community provides a 4bit quantized version of LingBot-World-Base (Cam), which reduces VRAM consumption and is intended for inference only; visual fidelity and temporal consistency may slightly decrease.
Usage The code is built on Wan2.2, and installation should refer to its documentation. The steps include: cloning the repository, installing dependencies, installing flash_attn, and downloading the model via huggingface-cli or modelscope-cli. Before inference, control signals need to be prepared (optional, can be generated from video using ViPE): intrinsics.npy (shape [num_frames, 4], values [fx, fy, cx, cy]) and poses.npy (shape [num_frames, 4, 4], transformation matrices in OpenCV coordinates). Multiple reference inference scripts are provided, run using torchrun, supporting 480P/720P resolutions, with the number of frames adjustable via --frame_num (e.g., 961 generates a one-minute video at 16 FPS); when VRAM is insufficient, --t5_cpu can be used to reduce memory. LingBot-World-Fast provides generate_fast.py and run_fast.sh scripts.
Timeline January 29, 2026: Released technical report, code, and models. March 2, 2026: Released LingBot-World-Base (Act) model weights. April 2, 2026: Released LingBot-World-Fast model weights. April 7, 2026: Released LingBot-World-Fast inference scripts. April 10, 2026: Released LingBot-World-Base (Act) user-friendly scripts.
Limitations The original repository is no longer actively maintained; the quantized model is intended for inference only and may have degraded quality; usage requires certain GPU resources.
License Apache 2.0 License.
