The Robbyant team announced the release of LingBot-Video on July 9, publishing the model under the Apache 2.0 license and providing a free demo link. The flagship variant packs 30 B total parameters but only activates 3 B experts per forward pass, a design that the authors say yields roughly three‑times faster inference compared with dense equivalents.
What the model delivers
LingBot‑Video is built on a mixture‑of‑experts (MoE) backbone, routing each token to a subset of specialized expert networks. In benchmark tests on RBench it scored 0.620, outpacing closed‑source competitors like Wan 2.6 (0.607) and Seedance 1.5 Pro (0.584). For quadruped motion synthesis it topped the chart with a 0.758 rating. The model was trained on over 70 000 hours of robotics footage, covering manipulation, navigation, and egocentric views, and incorporates a reward system that balances physical plausibility, task completion, aesthetics, and prompt adherence.
Two deployment options
- Dense 1.3 B – fits on consumer‑grade GPUs and is intended for quick prototyping.
- MoE 30 B‑A3 B – the full expert‑sparse model paired with a Refiner step that outputs 480 px video for higher visual quality. Both variants are available via the same repository, and the team provides a hosted demo for quick testing.
Caveats and cost considerations
While the MoE design reduces active parameters per inference, the total model size still requires substantial VRAM for the full 30 B variant (multiple A100‑class GPUs are recommended). The “~3× faster” claim holds when the routing overhead is amortized; on lower‑end hardware you may see little speed gain and higher memory pressure. Additionally, the open‑source license removes vendor lock‑in but leaves you responsible for maintenance, scaling, and any downstream bugs.
When to try it
If your startup is already experimenting with video generation and has access to mid‑range GPU clusters, spin up the 1.3 B dense model for proof‑of‑concept work. Reserve the 30 B MoE version for workloads where inference speed and quality justify the extra hardware cost. Keep an eye on community forks for potential optimizations and early‑stage tooling.