Ant Group unveils LingBot-VLA 2.0, an open model for controlling robots of any design

Robbyant, the robotics division of Ant Group, released LingBot-VLA 2.0 on July 8, 2026. It is an open vision-language-action model with 6 billion parameters for controlling robots of different designs, published under the Apache 2.0 license.

What the model was trained on

LingBot-VLA 2.0 was pretrained on roughly 60,000 hours of data, according to MarkTechPost. Of these, 50,000 hours are robot trajectories collected across 20 different robot configurations, and another 10,000 hours are egocentric first-person video of people performing everyday and manipulation tasks.

• Developer: Robbyant, an Ant Group division; released on July 8, 2026
• License: Apache 2.0, with the model checkpoint available for download
• Model size: 6 billion parameters (6B)
• Pretraining data: about 60,000 hours (50,000 hours of robot trajectories across 20 robot configurations + 10,000 hours of human video)
• Unified action space: 55 dimensions for all robot types

How the model controls different robots

LingBot-VLA 2.0 maps the control of any robot into a single canonical 55-dimensional action space that covers arms, dexterous hands, waist, head and mobile bases, MarkTechPost explains. This tackles the main problem of cross-embodiment models: previously, each robot design, with its own number of joints, gripper type or presence of a mobile base, required a separate control model. A unified action space lets the same checkpoint run on all 20 configurations.

A token-level action module built on Mixture-of-Experts helps scale the model without losing quality, and it needs no auxiliary load-balancing loss between experts. This makes it possible to grow model capacity without adding an extra loss term, which usually slows down and complicates training of such architectures.

Where the model gets geometry and timing

Dual-query distillation from the LingBot-Depth and DINO-Video models adds geometric and temporal supervision, which helps LingBot-VLA 2.0 account for the future state of the scene when choosing an action, according to MarkTechPost. In other words, the model "understands" in advance how the scene geometry and object positions will change after the current step, rather than reacting only to a static frame.

How far LingBot-VLA 2.0 outpaces its competitors

On the GM-100 test suite for general-purpose robot agents, LingBot-VLA 2.0 outperformed π0.5 and the previous version, LingBot-VLA-1.0, on both robot platforms tested, MarkTechPost reports. Beating both its predecessor and a third-party model on the same benchmark indicates that the improvements (the unified action space, the MoE module and dual-query distillation) deliver a measurable quality gain rather than simply adding parameters.

What this means

An open 6-billion-parameter model that performs equally well on 20 different robot configurations lowers the entry barrier for teams that want to train general-purpose robot agents without collecting a separate dataset for each specific robot design.

Frequently asked questions

What does VLA in the model's name stand for?

VLA stands for vision-language-action. The model processes an image and a text instruction and generates a robot action in a single pass.

Can LingBot-VLA 2.0 be used for free?

Yes. The model is released under the Apache 2.0 license, which permits free use, modification and commercial use of the checkpoint, including integration into your own robotics products.