On July 10, 2026, Ant Group’s LingBo division introduced LingBot-VA 2.0—the industry’s pioneering embodied virtual action model. Engineered from the ground up to address the core challenges of dynamic modeling, causal prediction, and real-time execution in robot-environment interactions, this model employs an autoregressive architecture trained entirely from scratch, avoiding dependency on capabilities transferred from digital world models. LingBot-VA 2.0 establishes its native framework through four strategic innovations:

• Integrating a semantic vision-action tokenizer as the visual encoder to enhance perception-action alignment.
• Adopting a strict causal pre-training paradigm to ensure temporal coherence and logical reasoning.
• Incorporating a Mixture-of-Experts (MoE) architecture to optimize performance-efficiency trade-offs.
• Achieving real-time closed-loop control via an enhanced asynchronous inference mechanism.

These advancements enabled the model to demonstrate exceptional performance in real-world robotic testing, achieving 150Hz real-time inference efficiency on a single GPU. Additionally, Ant Group’s LingBo released a suite of complementary models to collectively form an integrated embodied AI technology ecosystem.