Robbyant выпустила LingBot-World-Infinity — каузальную модель для интерактивных миров

On July 9, 2026, Robbyant, a division of Ant Group, introduced LingBot-World-Infinity — a causal video model with 14 billion parameters that functions as an interactive world simulator. The key feature: the model generates video in real time (720p, 60 frames per second) and responds to user commands.

What This Is and Why It Matters

LingBot-World-Infinity (version 2.0 of the LingBot-World series) — an interactive video model that predicts future video frames based on user actions. The model simultaneously processes text commands, virtual camera position, and physical character actions: shooting, spells, attacks, archery.

This solves two classic problems of video models: quality drift during long videos (textures blur, geometry distorts) and response latency (the model cannot react in real time).

• Main version: 14 billion parameters
• Lightweight version: 1.3 billion parameters, runs on a single GPU
• Performance: 720p video, 60 frames per second
• Interaction: two modes — direct semantic commands and object tracking via SAM

How It Solves Drift: MoBA Architecture

The core of the technology is MoBA-type attention (Mixture of Bidirectional and Autoregressive). The problem with standard autoregressive models: as context grows, the model begins to rely on it instead of predicting new frames, causing quality degradation and image blur.

MoBA adds a bidirectional attention block as a regularizer. This helps the model predict the future by relying on all available information, rather than just the nearest context.

The second component is two-stage training with distribution distillation (DMD). First, the model is trained on conditional stream generation, then the multi-step teacher is compressed into a fast version. The critical point: distillation optimization is applied to long trajectories of the model's own predictions, not just teacher examples. This forces the student to optimize on the same distribution of states it generates, preventing drift.

Interactive Mode: Director and Pilot

LingBot-World-Infinity operates within the Director-Pilot Co-Simulation Framework — a system of two interacting agents. The Director is a video-language model (VLM) that understands semantics, causality, and makes decisions. The Pilot is a Diffusion Transformer that simulates physics and renders video frames.

The system supports two interaction modes. In Direct Semantic Interaction mode, the VLM reads the current frame and generates events without requiring object masks. In Tracking-Assisted Object Interaction mode, the SAM model (Segment Anything Model) tracks objects of interest, the user selects an object and launches an action (for example, opens a door or rotates a ball).

Additionally, users can use text commands to change the global state of the scene (time of day, weather) and add objects. The VLM automatically selects semantically appropriate positions for their appearance.

What This Means

LingBot-World-Infinity demonstrates that interactive video simulators are transitioning from research labs into practical tools. Such systems are useful for game engine development, robotics training through simulation, and user interface design. Growing real-time performance makes them suitable for industrial applications.