LingBot Video: The First Open Source MoE Video Generation Model

Released July 9, 2026 under Apache 2.0, LingBot-Video is the first large scale open source video generation model built on Mixture-of-Experts architecture. It tops the RBench leaderboard among all open source models with an average score of 0.620, outperforming Cosmos3 Super, Wan 2.2 A14B, and HunyuanVideo 1.5, and beating several closed source commercial models including Google Veo 3.

What Is LingBot Video

LingBot-Video is a DiT based video generation model developed by the Robbyant team and released alongside its technical paper on arXiv (arXiv:2607.07675). The paper describes it as the first large scale, open source MoE video foundation model. It supports text-to-video, image-to-video, and text to image generation from a single unified architecture.

The model ships in two variants released under Apache 2.0. A Dense 1.3B parameter version is designed for consumer hardware inference. The MoE flagship uses 30 billion total parameters with 3 billion active per forward pass (30B-A3B), adds a Refiner step, and reaches the benchmark scores in the table below.

Both variants support the same core tasks. Text-to-video takes a prompt and generates a video clip. Image-to-video takes an image and a text prompt and animates the scene. Text to image produces a single frame. The LingBot World open source world simulator from the same team covered a different architecture focused on interactive 3D environments. LingBot-Video targets video generation directly.

Why MoE Changes the Equation

Dense transformer architectures activate every parameter during every inference step. A 30-billion parameter dense model requires proportionally more compute than a 3-billion parameter one. LingBot-Video replaces that with Mixture-of-Experts routing, where each token is sent to a subset of specialized expert networks rather than the full parameter set.

The practical result per the GitHub README: "balanced between capacity and cost with ~3x faster inference." The 30B-A3B designation means the model carries 30 billion parameters total but activates only 3 billion during any given forward pass. Inference compute stays closer to a 3B model while the model retains the representational range of 30B.

This is the same architectural principle behind the most capable language models. LingBot-Video applies it to video generation at open source scale, which no prior public video model had done before this release.

Trained on the Physical World

Standard video generation models learn from internet video, which over-represents stylized, edited, and non-physical content. LingBot-Video supplements that base with 70,000+ hours of robotics footage covering manipulation tasks, navigation scenarios, and egocentric perspectives recorded during physical interaction with real environments.

The training also introduces a multidimensional reward system that goes beyond standard criteria. In addition to aesthetics, prompt following, and motion consistency, the model is rewarded for physical rationality and task completion. This pushes generated motion toward real world physics rather than footage that merely looks plausible on the surface.

The connection to the benchmark results is direct. LingBot-Video's strongest scores on RBench come in categories where physical realism matters most, including Quadruped motion (0.758) and Manipulation (0.578).

General and Embodied Video Simulation

The clips below are generated by LingBot-Video from text prompts. This first group covers general scene generation and embodied AI scenarios, the category where the model's robotics training data has the most direct influence.

Material and Lighting Fidelity

The physical training data also improves the model's handling of material properties and lighting behavior. These clips demonstrate generation across reflective, transparent, and textured surfaces under varying light conditions.

Motion and Dynamics

Physical reasoning extends to dynamic systems. These clips demonstrate generation of motion driven by physical forces, including fluids, particles, and object interactions.

Benchmark Results

LingBot-Video was evaluated on the RBench leaderboard, which measures video generation quality across dimensions relevant to robotics. Categories include manipulation, spatial reasoning, multi entity scenes, long horizon consistency, physical reasoning, and motion types spanning single arm, dual arm, quadruped, and humanoid.

Model	Open source	Avg. score
LingBot-Video	Yes	0.620
Wan 2.6	No	0.607
Cosmos3 Super	Yes	0.581
Seedance 1.5 Pro	No	0.584
Veo 3	No	0.563
Wan 2.2 A14B	Yes	0.507
HunyuanVideo 1.5	Yes	0.460

Among open source models, LingBot-Video leads by 0.039 points over Cosmos3 Super. It also outscores Veo 3 (0.563) and Seedance 1.5 Pro (0.584), both closed source commercial products. Cosmos3 Super and HunyuanVideo 1.5 are both strong recent releases. LingBot-Video clearing them on physical reasoning metrics is the clearest evidence of what the robotics training data contributes.

The strongest individual category score is Quadruped motion at 0.758, where the model's exposure to footage of four legged robotic locomotion gives it a direct advantage. Wan 2.2 scores 0.507 overall on the same benchmark.

Two Variants and the Inference Pipeline

The Dense 1.3B variant handles text-to-video, image-to-video, and text to image on consumer GPUs. The MoE 30B-A3B variant adds the Refiner step for higher output quality and generates at up to 480p across all supported tasks.

Both variants use a two-step inference workflow. A prompt Rewriter based on Qwen3.6-27B converts plain text into a structured JSON caption that the DiT model processes. Users can paste a pre-formatted LingBot JSON caption to skip the rewriting step entirely. Multi-GPU inference is supported via FSDP and CP8 sharding. The Python installation requires torch 2.12 or newer.

The Rewriter is available as a LoRA adapter on top of the full Qwen3.6-27B base model. Both the adapter and a standalone version are published on HuggingFace alongside the main generation models.

Try It Now

You can generate text-to-video, image-to-video, and text to image directly in the community demo below, powered by the MoE 30B-A3B model at 480p. Plain text prompts are automatically expanded into structured JSON captions.

Text-to-video and image-to-video generation are also available in AI FILMS Studio using the latest open source models.

---

Sources

Project Page: technology.robbyant.com/lingbot-video GitHub: robbyant/lingbot-video HuggingFace: robbyant/lingbot-video-moe-30b-a3b HuggingFace (Dense): robbyant/lingbot-video-dense-1.3b HuggingFace Space: victor/lingbot-video arXiv: Scaling Mixture-of-Experts Video Pretraining for Embodied Intelligence