Today we open-source LingBot-Video — the first MoE-based video foundation model built for embodied intelligence. 🔹30B params, only 3B active at inference. 🔹Augmented with 70K hours of embodied data on top of large-scale internet video pretraining. 🔹Already outperforming Wan2.6, Seedance 1.5 Pro, and Cosmos3 Super on RBench. 🧵👇

What makes LingBot-Video built for embodied AI — robots need physical grounding, not just visual beauty. We redesigned the pretraining paradigm from scratch: ⚡ Architecture: DiT + MoE delivers ~3x inference efficiency over dense models — meeting the real-time demands of robot control loops. 📦 Data: 70K hours of embodied data (VLA, VLN, ego-centric) covering dexterous manipulation, navigation, and first-person interaction — teaching the model how actions change environments, not just how videos look. 🎯 Training: Multi-dimensional RL rewards align on physical plausibility and task completion, not just aesthetics.

On RBench (Peking University & ByteDance), LingBot-Video scores 0.620, surpassing Wan2.6 (0.607), Seedance 1.5 Pro (0.584), and Cosmos3 Super (0.581) in generating physically grounded robot behaviors. LingBot-Video is now fully open-source. 🔗 Model: 🔗 Code: 🔗 Paper: 🔗 Project Page:

You're on a roll lately!

Other video models generate pretty frames. LingBot-Video generates physically grounded behaviors. Huge difference. Open-sourcing this is a big deal 👏

Love seeing more open-source progress in embodied AI. Great release!

The use of MoE stands out to me. 3B active out of 30B is promising for long-context embodied workloads where inference cost is a real bottleneck. Curious how the physical-reasoning rewards hold up on downstream robotics tasks. Nice release.

Reinforcement learning improving physical consistency instead of just aesthetics is an interesting shift. Curious to see how it performs in robotics tasks.

To use for simulating and generating videos to train robots?

This open-source release is a big win for embodied AI. The focus on physical grounding over pretty visuals is exactly what's needed!!

This is much more interesting than another model chasing visual realism.

This feels like the right direction for embodied AI. Prioritizing physical reasoning over visual polish, backed by MoE and 70K hours of VLA/VLN data.

how can I give this a try?

I'd be interested in downstream robotics evaluations

The sparse MoE design is one of the most interesting parts of this release

Physical reasoning deserves more attention.

@NeuraFlowAix Looking forward to community experiments.

Great move for the embodied AI community. LingBot Video brings open source video generation closer to robotics where motion, contact and physical consistency matter more than just beautiful frames.

70K+ hours of embodied data gives this a very different training signal.

70k+ hours of embodied data is a huge advantage for LINGBOT-VIDEO. Training at this scale gives the model a much stronger understanding of real-world interactions, making it about far more than just generating videos. Excited to see how this pushes embodied AI forward.

Physical reasoning is quickly becoming its own benchmark.

The 30B/3B sparse MoE setup is the biggest technical takeaway.

The 70K hours of embodied data completely changes how I look at this release. That’s a serious investment in real-world intelligence.

The shift from optimizing visual realism to optimizing physical understanding is what stands out most. Efficient MoE inference paired with large-scale embodied pretraining could be a meaningful step forward for robotics and embodied AI.

Open source should make it much easier for the research community to evaluate and improve the model.

Making LINGBOT-VIDEO open source is a great move. It gives the community the opportunity to explore, test, and build on the model, helping accelerate innovation and uncover its full potential.

RL for physical plausibility is one of the strongest parts of this release.

Curious to see how this performs on real robotics tasks beyond benchmarks.

Excellent content! Looking forward to seeing more posts like this in the future.

Innovative and outperforming embodying intelligence

The idea of a world that keeps evolving instead of waiting for your next prompt is a really cool direction for interactive AI.

This looks incredible! 🔥 Open-sourcing LingBot-Video with that MoE efficiency and heavy embodied data focus is a big win for robotics and real-world AI. Can't wait to see what the community builds with it. Great work Robbyant team! 🚀

@OtakuMachine Gostei bastante do foco em raciocínio físico.

30b params but only 3b active is nuts, MoE really making these video models way more efficient.

The setup for real-time inference is sooo smart. Robots actually need that kind of efficiency.

@EmmaUsesAi Smart engineering choices throughout

LingBot-Video is setting a new benchmark for open-source embodied AI.