Ant Lingbo Open-Sources LingBot-Video, the World's First MoE Video Base Model for Embodied Intelligence

Ant Lingbo released LingBot-Video as an open-source model. The 30-billion-parameter mixture-of-experts system targets video generation for embodied intelligence tasks. It activates roughly 3 billion parameters during inference.

The release introduces a model trained on 70,000 hours of robot-related video data. Those hours cover VLA, VLN, and Ego datasets. A multi-dimensional reinforcement learning reward system aligns outputs with physical constraints and task goals.

LingBot-Video scored 0.620 on RBench and placed first on Physics-IQ Verified. The numbers exceed Wan 2.6 and other listed dense models. Efficiency claims state a threefold gain over same-scale dense architectures.

Model Architecture Details

The mixture-of-experts design routes each token through a subset of experts. Only 3 billion parameters run per forward pass. This approach reduces memory and compute compared with full 30-billion-parameter activation.

Training incorporated robot trajectories, navigation sequences, and first-person recordings. The reward model scores physical plausibility separately from task completion. Both scores guide updates during reinforcement learning stages.

The model supports downstream uses such as action prediction and synthetic data creation for robot training. Developers can fine-tune on narrower robot domains without retraining the full base.

Benchmark Performance Breakdown

RBench measures video quality, temporal consistency, and task relevance. LingBot-Video posted the 0.620 aggregate. Individual category scores were not released in the initial announcement.

Physics-IQ Verified focuses on adherence to Newtonian rules in generated clips. The model ranked first among listed submissions. Organizers noted fewer violations of object permanence and collision dynamics than prior entries.

These results come from the company’s published evaluation tables. Independent replication has not yet appeared in peer-reviewed venues.

Industry Comparison Points

Wan 2.6 operates as a dense transformer at comparable total size. Its inference cost remains higher because all parameters stay active. LingBot-Video’s selective activation changes that cost structure.

Other embodied video efforts still rely on smaller dense models or diffusion pipelines. Those pipelines often require many denoising steps. The new MoE design targets fewer total FLOPs per second of generated video.

Existing open robot datasets remain smaller than the 70,000-hour corpus cited. Scale alone does not guarantee quality, yet the volume allows broader coverage of edge cases.

Use Cases in Robotics Pipelines

Robot action prediction benefits from forward simulation of camera views. LingBot-Video can generate candidate next frames given current observation and planned joint commands. Engineers can rank those frames before sending commands to hardware.

Simulation data generation supplies additional training examples when real robot time is limited. Synthetic clips that respect physics can expand datasets for policy learning. Early internal tests showed reduced sim-to-real gap on simple grasping tasks.

Navigation research can use ego-centric video rollouts to test path planning without physical robots. The model supplies visual feedback consistent with the simulated environment layout.

Limitations and Open Questions

The reward system relies on automated scoring functions. Those functions may miss subtle failure modes that human operators notice. Long-horizon tasks still show drift in generated sequences.

No public license details beyond the open-source announcement have been confirmed. Downstream commercial use terms remain to be clarified by the maintainers.

Independent verification of the 70,000-hour dataset composition has not occurred. Data provenance directly affects bias and safety properties for any deployed system.

Next Signals to Monitor

Watch for third-party benchmark reproductions on public leaderboards within the next three months. Consistent RBench gains would strengthen the efficiency claims. Divergent results would highlight evaluation gaps.

Track adoption metrics on major robot learning repositories. Download counts and fine-tune forks offer early usage signals. Sustained activity over six months would indicate practical utility.

Observe whether competing labs release similar MoE video models. A rapid response from dense-model teams would show how the efficiency trade-off is valued across the field.