LingBot-Video: First Open-Source MoE Video Foundation Model for Embodied Intelligence
Ant Group's Robbyant open-sources LingBot-Video, a 30B-A3B MoE video generation model achieving SOTA on RBench with Apache 2.0 license.
Ant Group's embodied AI subsidiary Robbyant (Ant Lingbo) has open-sourced LingBot-Video, the first large-scale Mixture-of-Experts (MoE) video foundation model purpose-built for embodied intelligence. The release includes both a 30B-A3B MoE model and a 1.3B dense model, along with code and a prompt rewriter, all under the Apache 2.0 license.
Architecture and Design
LingBot-Video uses a DiT-based architecture with sparse MoE, activating only 3B parameters per token while maintaining 30B total capacity. This design achieves ~3x faster inference than dense models of similar capability. The model supports three generation modes:
- Text-to-Image (T2I)
- Text-to-Video (T2V)
- Text-Image-to-Video (TI2V)
A separate refiner model is included for quality enhancement, and a prompt rewriter based on Qwen3.6-27B converts plain text prompts into structured JSON captions.
Training Data and Reward System
The model was trained on massive web video data integrated with 70,000+ hours of embodied data, covering physical motion, human action, robot tasks, and open-ended creation. A multi-reward system optimizes for three dimensions:
- High aesthetics — visual quality
- Physical rationality — adherence to real-world physics
- Task completion — usefulness for embodied AI tasks
Benchmark Performance
On the RBench leaderboard, LingBot-Video ranks #1 among all open-source models, surpassing Cosmos3 Super, LongCat-Video, Wan 2.2 A14B, and HunyuanVideo 1.5. It achieves an average score of 0.620, particularly excelling in manipulation tasks (0.578), long-horizon tasks (0.634), and quadruped scenarios (0.758).
The model also outperforms several closed-source systems including Wan 2.6, Seedance 1.5 Pro, and Veo 3 in multiple categories.
ComfyUI Integration
ComfyUI creator comfyanonymous has already opened a WIP pull request for native LingBot-Video support, indicating direct ComfyUI integration is on the way.
Companion Release: LingBot-World 2.0
Alongside LingBot-Video, Robbyant also unveiled LingBot-World 2.0, a world model capable of hour-long real-time generation. This marks a significant step toward bridging digital intelligence with physical world understanding.
Availability
The models are available on Hugging Face:
- MoE (30B-A3B): robbyant/lingbot-video-moe-30b-a3b
- Dense (1.3B): robbyant/lingbot-video-dense-1.3b
- Rewriter LoRA: robbyant/lingbot-video-rewriter-lora
Source code and inference scripts: github.com/Robbyant/lingbot-video
Research paper: arXiv:2607.07675