What Changed
Robbyant has introduced LingBot-Video, an open-source large-scale Mixture-of-Experts (MoE) video generation model. This release marks a significant step towards integrating video synthesis with embodied intelligence, focusing on generating videos that reflect physical world understanding. The project includes the technical report, code, models, and rewriters, all released under an Apache 2.0 License.
LingBot-Video differentiates itself through its MoE architecture, which is designed for efficiency and scalability. It has been trained on a substantial dataset comprising massive web videos combined with over 70,000 hours of embodied data. The model's training incorporates a multi-reward system, optimizing for high aesthetic quality, physical rationality, and task completion within generated videos.
Technical Details
LingBot-Video leverages an efficient MoE architecture, enabling approximately 3x faster inference compared to dense models while maintaining capacity. The model suite includes several components:
- LingBot-Video-Dense (1.3B parameters): A dense model for Text-to-Image (T2I), Text-to-Video (T2V), and Image-to-Video (TI2V) tasks.
- LingBot-Video-MoE (30B-A3B parameters) + Refiner: The primary MoE model, supporting T2I, T2V, TI2V, and refinement capabilities.
- LingBot-Video-Rewriter-Base (Qwen3.6-27B official): A prompt rewriter for expanding user prompts.
- LingBot-Video-Rewriter-Adapter (Qwen3.6-27B LoRA): A prompt rewriter specifically for JSON output.
The recommended inference workflow involves a three-stage process: prompt rewriting, automatic negative prompt generation, and unified inference. The prompt rewriter converts plain natural-language prompts into structured JSON captions. An Auto Negative block then prunes the negative prompt based on this caption. Finally, the unified inference runner executes the video generation, supporting both direct diffusers and SGLang Diffusion backends. For multi-GPU inference, the --enable_fsdp_inference flag shards the base DiT and refiner DiT across GPUs, reducing memory pressure, though sufficient system RAM is required for loading large MoE checkpoints.
The project provides installation instructions and a requirements.txt file specifying dependencies like torch (recommended 2.12.0.dev20260220+cu130), transformers (5.8.1), and diffusers (0.39.0). Optional SGLang dependencies are available for SGLang Diffusion or fused/FP8 MoE runtime. Ready-to-run scripts are provided for single-GPU and multi-GPU inference configurations, covering dense and MoE models for T2I, T2V, and TI2V tasks, with options for refiner integration and FSDP/CP8 optimizations.
Benchmark Analysis
As of July 9th, 2026, LingBot-Video holds the top rank on the RBench Leaderboard. The model achieved an average score of 0.620. Specific category scores include:
- Manipulation: 0.578
- Spatial: 0.643
- Multi-entity: 0.444
- Long-horizon: 0.634
- Reasoning: 0.505
- Single arm: 0.636
- Dual arm: 0.639
- Quadruped: 0.758
- Humanoid: 0.689
Compared to other open-source models like Cosmos3 Super (Avg. 0.581), LongCat-Video (Avg. 0.437), Wan 2.2 A14B (Avg. 0.507), and HunyuanVideo 1.5 (Avg. 0.460), LingBot-Video demonstrates superior overall performance. Against closed-source models such as Wan 2.6 (Avg. 0.607), Seedance 1.5 pro (Avg. 0.584), and Veo 3 (Avg. 0.563), LingBot-Video maintains a competitive edge, particularly in average score and several specific categories like Manipulation, Long-horizon, and Quadruped tasks.
Developer Implications
Developers working on embodied AI, robotics, or advanced video generation now have access to an open-source MoE model that prioritizes physical rationality and task completion. The Apache 2.0 License allows for broad use and modification. The provided inference workflow, including prompt rewriting and auto-negative generation, offers a structured approach to leveraging the model effectively. The support for both diffusers and SGLang backends, along with multi-GPU inference scripts, provides flexibility for deployment across different hardware configurations.
The emphasis on structured JSON captions for inference suggests that developers will need to integrate the prompt rewriter into their pipelines or design their prompt inputs accordingly. The model's ability to generate physically rational videos could be particularly valuable for simulating complex robotic tasks or creating realistic training data for embodied agents. The efficiency gains from the MoE architecture could also reduce computational costs for certain applications.
Bottom Line
LingBot-Video represents a notable advancement in open-source video generation, specifically tailored for embodied intelligence. Its MoE architecture, trained on extensive embodied data, delivers efficient inference and strong performance in generating physically coherent and task-oriented videos. The comprehensive release, including code, models, and detailed inference workflows, provides developers with a powerful tool for applications requiring sophisticated video synthesis with a focus on real-world understanding.
Top comments (0)