Independent guide — Lingbot is not affiliated with Robbyant or Ant Group · Last verified: July 12, 2026
The LingBot-Video model, explained
LingBot-Video is Robbyant’s (Ant Group) open-source video generation model — the first large-scale Mixture-of-Experts video model dedicated to embodied intelligence. This guide tracks what is actually released, how it benchmarks, and how to use it.
Availability status
| Item | Status | Notes |
|---|---|---|
| Code | ✅ Released | GitHub · Robbyant/lingbot-video |
| Weights | ✅ Released | Hugging Face & ModelScope (Dense 1.3B, MoE 30B-A3B) |
| Hosted API | ❌ Not available | Not on fal.ai or Replicate as of our last check |
| ComfyUI | ⏳ No official support | Community interest exists; nothing official yet |
| Commercial use | ✅ Allowed | Apache 2.0 license |
What it is — and what it isn’t
If you arrive expecting another Sora-style cinematic generator, recalibrate: Robbyant’s own launch line was “robots don’t need beautiful videos.” LingBot-Video is a DiT-based video pretraining paradigm for embodied intelligence (arXiv:2607.07675). It exists to give robots a visual imagination that respects physics.
- Efficient MoE architecture — 30B parameters with only 3B active per step, scaled from scratch, roughly 3× faster inference than dense counterparts.
- Embodied data engine — massive web video plus 70,000+ hours of embodied data: manipulation, egocentric views, navigation.
- Multi-reward training — rewarded jointly for aesthetics, physical rationality, and task completion.
- Three tasks in one model — T2V, T2I, and TI2V, plus an optional refiner in the MoE package.
Dense 1.3B vs MoE 30B-A3B
| Dense 1.3B | MoE 30B-A3B | |
|---|---|---|
| Tasks | T2I, T2V, TI2V | T2I, T2V, TI2V + Refiner |
| Hardware | Single GPU (official scripts) | Multi-GPU FSDP + CP8, large host RAM |
| Best for | Trying the model, small setups | Maximum quality, research rigs |
RBench benchmark (official README, July 9, 2026)
RBench evaluates embodied video generation: manipulation, spatial reasoning, multi-entity scenes, long-horizon consistency, and reasoning. Average scores:
| Model | Access | Avg |
|---|---|---|
| LingBot-Video | Open | 0.620 |
| Wan 2.6 | Closed | 0.607 |
| Seedance 1.5 pro | Closed | 0.584 |
| Cosmos3 Super | Open | 0.581 |
| Veo 3 | Closed | 0.563 |
| Wan 2.2 A14B | Open | 0.507 |
| HunyuanVideo 1.5 | Open | 0.460 |
| LongCat-Video | Open | 0.437 |
Two takeaways: among open models LingBot-Video leads by a wide margin, and it also edges out closed models — Wan 2.6, Seedance 1.5 pro, and Veo 3 — which is rare for an open release in this category.
Downloads & official resources
- GitHub repository — code, scripts, docs
- Hugging Face collection — Dense 1.3B, MoE 30B-A3B, prompt rewriter weights
- Technical report — “Scaling Mixture-of-Experts Video Pretraining for Embodied Intelligence”
- Official project page
Using the model
Two things trip almost everyone up. First, the model does not take plain-English prompts — inference consumes structured JSON captions produced by a two-stage rewriter. Read the prompt guide before judging output quality. Second, hardware requirements are real: start with how to run it locally for the honest version-requirements table.
Want to create right now without a GPU? Our own Lingbot Video generator is free to try after sign-up — it uses different models (clearly labeled), not LingBot-Video.
Frequently asked questions
What is LingBot-Video?
LingBot-Video is the first open-source large-scale Mixture-of-Experts (MoE) video generation model built for embodied intelligence, released by Robbyant (Ant Group) on July 9, 2026. It supports text-to-video, text-to-image, and text+image-to-video, and prioritizes physical realism over pure aesthetics.
Is LingBot-Video free for commercial use?
Yes. LingBot-Video is released under the Apache 2.0 license, which permits commercial use, modification, and redistribution. Check the LICENSE file in the official repository for full terms.
Can I try LingBot-Video online?
Not yet. As of our last check there is no hosted API on fal.ai or Replicate, so running it means downloading the weights and using your own GPUs. Subscribe below and we will email you the moment a hosted option appears.
How is LingBot-Video different from Sora, Veo, or Wan?
Those models optimize for visual fidelity and creative content. LingBot-Video optimizes for physical rationality — how objects, hands, and robots actually move — because its target use is embodied AI. On the RBench embodied-video benchmark it leads with a 0.620 average.
What GPU do I need to run it?
The 1.3B Dense model is the realistic single-GPU option. The 30B-A3B MoE model requires multi-GPU FSDP sharding and substantial host memory. See our how-to-run guide for the honest hardware picture.