On July 11, 2026, Robbyant, the embodied-AI unit inside Chinese fintech giant Ant Group, unveiled LingBot-VA 2.0, a robot-manipulation model it bills as the first "embodied-native" foundation model. The claim is about method: rather than adapting an off-the-shelf video generator, Robbyant says it pretrained a causal diffusion transformer from scratch specifically for physical control.
Why 'from scratch'
The company argues that fine-tuning a video model for robotics carries baggage — visual latents that preserve appearance but little physical structure, denoising too slow for closed-loop control, and training objectives that never teach how actions reshape the world. Version 2.0's answer, in Robbyant's phrasing, is that it "pretrains a causal DiT natively," building the whole stack for embodiment instead of borrowing one from content generation.
The numbers
LingBot-VA 2.0 uses about 15.3 billion total parameters with roughly 2.5 billion active per token, split into a mixture-of-experts "video expert" (128 routed experts) and a dense "action expert." Robbyant reports 93.6% average success across 50 tasks on the RoboTwin 2.0 simulation benchmark — ahead of its own v1 at 92.2% and a prior method, Motus, at 87.9%. A stack of inference optimizations cut latency from 927 to 142 milliseconds per chunk, enabling asynchronous control at 225 Hz, roughly a sixfold speedup.
The caveats
Every figure here is vendor-reported, drawn from Robbyant's own paper and project page and summarized by the outlet MarkTechPost, with no independent replication. The benchmark results are simulation-only; the model's real-robot performance is described qualitatively (tasks like air hockey and conveyor-belt handling). And unlike the sibling LingBot-VLA 2.0 — a separate 6-billion-parameter model released under Apache-2.0 on July 8 — open weights or code for VA 2.0 have not been confirmed released, so it is best read as a paper-and-page announcement for now.
The China embodied-AI race
The launch caps an unusually dense week: Robbyant shipped depth, vision, a vision-language-action policy and a world model within days of one another. Taken together, the cadence signals Ant Group scaling a full embodied-AI stack, and it slots into a broader Chinese push into robotics foundation models and open releases.