Robbyant, the embodied AI unit inside Ant Group, has released the LingBot-VA 2.0, the first embodied-native foundation model described as a video-action foundation model for generalist robot manipulation. The research team pretrains the whole stack for embodiment instead of fine-tuning a video generator. Version 2.0 pretrains a causal DiT natively, replacing the approach of Version 1.0 which finetuned a stack into a causal model. The tokenizer replaces the compression-only VAE, adding semantic alignment and a latent-action objective; world states and actions now share one latent space. Version 2 pretrains a causal DiT with a sparse MoE video stream, keeping the Mixture-of-Transformers layout. The video expert replaces its dense FFN with a sparse MoE routed layer holding 128 routed SwiGLU experts, top-8 routing, and one shared expert. The video backbone is roughly 13.0B parameters, about 1.9B active; with the action expert and MCP heads, training covers about 15.3B parameters, with roughly 2.5B activating per token at inference. Training uses a rectified-flow objective with a hybrid Muon plus AdamW optimizer. Multi-chunk prediction (MCP) fixes myopic supervision; in ablation it matched the baseline’s 45k-step accuracy in 20k steps, a 2.