LingBot-Video - AntLingbo Open Source Video Model for Embodied Intelligence - AiBoss

What is LingBot-Video?

LingBot-Video is the world's first open-source MoE video generation model for embodied intelligence, developed by AntLingbo Technology. Based on a DiT + MoE architecture, the model has a total of 30B parameters, with only about 3B activated, balancing large capacity and efficient inference. Integrating 70,000 hours of internet and robotics data, the model uses multi-dimensional reinforcement learning rewards to align physical plausibility with task completion, ranking first in both RBench and Physics-IQ benchmarks. It can generate videos of robot operations and interactions that conform to real-world physical laws.

Main functions of LingBot-Video

Text/Image to Video GenerationSupports unified generation of T2V, TI2V, and T2I, covering open-world scenarios and embodied tasks.
Robot motion predictionGenerates motion sequences and contact states for robotic arms, humanoid robots, quadruped robots, etc.
Physical simulation video generationSimulates real physical phenomena such as material deformation, fluid motion, and light reflection.
Simulation data synthesis: Provides large-scale, low-cost, physically appropriate video data for robot strategy training.
World Model ResearchAs a physics simulator, it supports strategy evaluation and action planning.

The technical principle of LingBot-Video

DiT + Sparse MoE architectureThe model employs a Single-Stream Diffusion Transformer to uniformly process visual latent codes and conditional tokens, replacing Dense FFN with MoE. Of the total 30B parameters, only approximately 3B are activated, achieving about 3x inference speedup. 128 fine-grained routed experts, in conjunction with shared experts, capture different physical patterns through top-K routing, avoiding interference from subtasks.
Data profiling engine and 70,000 hours of embodied dataBased on massive amounts of internet video, robot data such as VLA, VLN, and Ego are introduced, covering dexterous manipulation, navigation, and first-person perspective interaction, enabling the model to learn the intrinsic relationship between actions and environmental changes.
Multidimensional reinforcement learning reward alignmentIn addition to aesthetics, prompt following, and motion consistency, the training phase introduces physical plausibility and task completion rewards, and uses real-world videos as preference signals to ensure that the generated results conform to real physical laws and robot task requirements.
Unified Input and 3D RoPEThe T2I/T2V/TI2V tasks are unified into single-sequence tokens. The conditional tokens and visual tokens are mapped to non-overlapping spatiotemporal coordinates through multimodal 3D MM-RoPE, eliminating task-specific architectural requirements while maintaining spatial locality and temporal order.
Cascade generation and stable trainingThe model employs a cascaded design of a base generator and a refiner. The base generator generates compact latent codes, while the refiner enhances details. Combined with QK-Norm, AdaLN-Single modulation, and a sequence-level load balancing strategy without auxiliary loss, it ensures stable training of long sequences and high resolutions.

How to use LingBot-Video

Environmental preparationEnsure that Python 3.10+ and CUDA environments are configured locally or on the server, and that sufficient video memory is reserved.
Cloning repositoryPull the LingBot-Video official code repository from GitHub to your local working directory.
Install dependenciesInstall the required deep learning and inference libraries according to the project's requirements.txt file.
Download weight: Obtain the model checkpoint file of the corresponding size from HuggingFace or ModelScope.
Preparation promptsOrganize text descriptions, reference images, or structured control signals into an input format supported by the model.
Running inferenceLoad the model and execute the generation script to output embodied intelligence or open-world video.

LingBot-Video's core advantages

The world's first open-source embodied video MoE model: Fills the gap in open-source video infrastructure models for robots.
Decoupling capacity and efficiencyWith a parameter scale of 30B, only 3B are activated, resulting in inference speed approximately 3 times faster than a similarly sized Dense architecture, thus meeting real-time interaction requirements.
Physical rationality is leadingIt ranked first in both RBench (0.620) and Physics-IQ Verified, surpassing Wan2.6, Seedance 1.5 Pro, Cosmos 3, and others.
Data specialization70,000 hours of robot data injection enabled the model to understand action-environment causality, rather than just visual style.
Unified task frameworkT2I/T2V/TI2V single-model unified processing reduces the cost of multi-task deployment.

LingBot-Video's project address

Project official websitehttps://technology.robbyant.com/lingbot-video
GitHub repositoryhttps://github.com/Robbyant/lingbot-video
HuggingFace model libraryhttps://huggingface.co/collections/robbyant/lingbot-video
arXiv technical paper: https://arxiv.org/pdf/2607.07675

Comparison of LingBot-Video with similar competing products

Dimension	LingBot-Video	NVIDIA Cosmos 3 Super
Research and Development Background	Robbyant, an open-source technology focused on embodied intelligent video generation, has released its open-source code.	Released by NVIDIA in June 2026, this is a full-modal world model for training physical AI and robots.
Architecture Design	Single-Stream Sparse MoE based on DiT provides unified processing of T2I/T2V/TI2V.	The Mixture-of-Transformers (MoT) dual-tower architecture uses an autoregressive Transformer to process text and a diffusion Transformer to process continuous modalities.
Parameter size	Total parameters 30B, with approximately 3B active, achieving decoupling of capacity and efficiency.	Total parameters 64B (Super), Nano version is 16B, all parameters activated.
Modal support	Text, images, and videos; focusing on visual generation and motion condition modeling.	Text, images, videos, audio, and action sequences are processed uniformly, supporting full-modal input and output.
Training data	70,000 hours of internet video data, plus VLA/VLN/Ego robot data, are used to inject dynamic prior knowledge of actions and the environment.	20 trillion tokens, containing approximately 1 billion images, 400 million real and synthetic videos and environmental audio, covering scenarios such as robotics, driving, and warehousing.
Open source level	The model weights, code, and technical reports are all open source and can be freely downloaded and used for secondary development by the community.	The model weights and code are open source under the OpenMDW 1.1 license, allowing commercial use, and come with a synthetic dataset and benchmark.
RBench performance	With a score of 0.620, it ranks first in publicly available comparisons.	With a score of 0.581, it ranks among the top open-source models in some evaluations.
Physics-IQ	It ranked number one in the Physics-IQ Verified assessment.	With I2V scores of 43.8 and V2V scores of 59.7, they achieved state-of-the-art (SOTA) rankings in their respective leaderboards.
Reasoning efficiency	With a total of 30B parameters, only 3B are activated, resulting in inference speeds approximately three times faster than a similarly sized Dense architecture.	Full parameter activation for 64B requires H100-level hardware support; the Nano version is optimized for faster inference.

Application Scenarios of LingBot-Video

Robot simulation trainingGenerate physically plausible operation videos for robotic arms and humanoid robots, replacing expensive real data collection.
Simulation data synthesisGenerate diverse scenes and action trajectories in batches for policy pre-training and data augmentation.
World Model ResearchAs a predictor of the physical world, it supports imagination-based robot planning and decision-making.
Autonomous driving simulationGenerate road interaction and ego-perspective videos that conform to physical laws for use in verifying perception algorithms.
Service and ManufacturingSimulate human-computer interaction processes in warehousing, healthcare, and home service scenarios to optimize task strategies.