LingBot-World 2.0, 14B Open Source Released

On July 9, 2026, the Robbyant team released LingBot-World 2.0 (LingBot-World-Infinity), an open-source interactive world model with 14B parameters. By combining Causal pre-training and Distillation techniques, the model addresses the error accumulation and speed degradation issues faced by existing models, enabling real-time video generation at 720p resolution and 60fps.

Key technical features include:

• Infinite Interaction Segments: Overcomes the limitations of autoregressive methods to support long-sequence generation over extended time periods.
• Two-Stage Distillation: Applies Consistency Distillation and Distribution Matching Distillation (DMD) to simultaneously secure real-time response speed and image quality.
• Agentic Harness: Separates high-level semantic reasoning and low-level physical simulation by dividing the VLM into a 'Director' and the video generator into a 'Pilot'.

The deployment environment is based on the Wan2.2 codebase and requires PyTorch 2.4.0 or higher and flash-attention. Currently, the 14B causal-fast model weights and inference code have been released, while a 1.3B model for single-GPU deployment is scheduled for future release. For direct service deployment, refer to the SGLang recipe or NVIDIA flashdreams.