LingBot Launches 30B-Parameter Open Source Video Model Trained on 70,000 Hours of Robot Footage

aiai-modelingai-open-modelsai-model-releasesai-research-evals 6 posts · 6 accounts

LingBot on July 10 launched LingBot-Video, a sparse mixture-of-experts video model for embodied artificial intelligence that combines general internet footage with 70,000 hours of robot navigation data. The Apache 2.0-licensed release offers 30B and 3B parameter versions and ranked first on the RBench benchmark, outperforming several closed generative video competitors.

Rather than focusing primarily on visual aesthetics, the system is trained to understand physical cause-and-effect and motion by applying reward signals for task success. The weights and training harness are fully open source and available on Hugging Face for developers building robotics and autonomous systems.

From the sources (6 posts)

@omarsar0

Cool open-source release. LingBot-World 2.0 holds 720p at 60 fps in real time and stays coherent for a full hour of interaction. A 1.3B variant runs on a single consumer GPU, so you can actually run a world model at home. Fully open sour

@dair_ai

RT @omarsar0: Cool open-source release. LingBot-World 2.0 holds 720p at 60 fps in real time and stays coherent for a full hour of interact…

@solennemoon

Interactive world models finally got a major upgrade. LingBot-World 2.0 from Robbyant solves the drift issue that ruins most simulations. 🧠 Brain (VLM) + 🧠 Cerebellum (Gen) = 60 minutes of consistent, high-fidelity world building. https

@askalphaxiv

A video model that's built for robots! As most video models learn appearance, this paper LingBot-Video tries to learn action, motion, and physical cause-and-effect. It does this by scaling a sparse MoE video diffusion model, mixing intern

@adinayakup

LingBot-Video 🎬 MoE video model built for embodied AI from @robbyant_brain - 30B/3B - Apache 2.0 - Trained on web videos + 70K hours of embodied data - Tops RBench: ahead of Cosmos3/Veo 3/Seedance 1.5 pro

@victormustar

RT @AdinaYakup: LingBot-Video 🎬 MoE video model built for embodied AI from @robbyant_brain - 30B/3B - Apache 2.0 - Trained on web video…

Market impact

8 no move no near-term move expected

Why it matters: A niche open-source embodied-AI video/world model from a small lab (not a frontier general-purpose LLM) tops a robotics benchmark ahead of Nvidia's Cosmos and Google's Veo, but there is no public pure-play and no near-term earnings read-through, so it is a specialist research launch rather than an index- or thesis-moving DeepSeek-style event.

At publish time: Headline is an open-weight (Apache 2.0) model release from an obscure lab (LingBot/Robbyant) beating Cosmos3/Veo3/Seedance on RBench; frontier-open framing but confined to embodied-AI/robotics with no mega-cap earnings hook.

model confidence: medium

What happened next

1–2 weeks later

Not checked yet — this story is too recent.

Follow-up coverage

No follow-up coverage matched this story.

Fed's Waller Says Inflation Target Should Be a Range SK Hynix Raises $28.1 Billion in US Nasdaq Debut