Robbyant, an embodied AI company within Ant Group , is releasing LingBot-World 2.0 (Infinity), an open-source interactive world model that generates playable video environments in real time. It targets researchers and developers working on world models, game AI, and robotics simulation who need scenes that persist and react rather than short clips that end after a few seconds. The launch matters now because most publicly available generative video systems still cap out at brief, non-interactive outputs, while this release pushes toward hour-long continuous generation with 720p/60fps real-time output and a richer action vocabulary.
Under the hood, the model uses a Causal Pretraining Paradigm combined with a proprietary MoBA (Mask of Bidirectional Attention) mechanism, which lets it learn world evolution in chronological order and reduce the compounding errors that cause texture blurring, geometry collapse, and scene breakdown in long-horizon generation . On the interaction side, users can drive characters with keyboard input to attack, shoot arrows, cast spells, jump, and glide, and can trigger global events like day-night cycles and weather changes through text commands. A dual-agent setup pairs a Pilot Agent that plans and executes character behaviors with a Director Agent that introduces new events as the scene progresses , and the system supports multiple users within a single persistent world for collaborative exploration .
The release ships with a distilled fast inference version, an optimized generation pipeline, and day-0 support for SGLang , available on GitHub, Hugging Face, and the Reactor platform for browser trials. It arrives alongside LingBot-Video, described as an open-source video generation foundation model built on a Mixture-of-Experts architecture designed for embodied intelligence , signaling how Robbyant plans to connect generative world models to its broader work on foundational large models for embodied AI and robotic companions for use cases like elderly care, medical assistance, and household tasks .
Comments (7)
Sign in to join the discussion.
N
Nadia KovalenkoSep 10, 2026
One brain for all robots is a bold pitch when my Roomba still gets stuck on the same sock every Tuesday. Show me it navigating my hallway before it summons storms.
T
Tomi AdigunSep 10, 2026
The tweet buries the lede under three emojis and a truncated link. If the world runs for hours, lead with the hours, not the spellcasting.
V
Vikram RamaswamySep 10, 2026
Pre-recorded demo or live? Because 'runs for hours' and 'responds to every move' are two very different infra bills.
V
Vikram RamaswamySep 10, 2026
Also, 'open-source' doing heavy lifting here. Weights, inference code, or just a readme with a discord invite?
M
Min-Jae HwangSep 10, 2026
Cuts in the demo reel are way too fast to trust. When a world model is actually persistent you slow down and let one shot breathe for 20 seconds.
L
Lior Ben-EzraSep 10, 2026
Every generation gets the Minecraft it deserves. Ours hallucinates.
P
Prachi DevalSep 10, 2026
Curious how you're framing the leap from world model to robot foundation model, those are pretty different loss landscapes. The tagline and the demo feel like two separate launches stapled together.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Robbyant
Building One Brain for All Robots
Building One Brain for All Robots
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Robbyant, an embodied AI company within Ant Group , is releasing LingBot-World 2.0 (Infinity), an open-source interactive world model that generates playable video environments in real time. It targets researchers and developers working on world models, game AI, and robotics simulation who need scenes that persist and react rather than short clips that end after a few seconds. The launch matters now because most publicly available generative video systems still cap out at brief, non-interactive outputs, while this release pushes toward hour-long continuous generation with 720p/60fps real-time output and a richer action vocabulary. Under the hood, the model uses a Causal Pretraining Paradigm combined with a proprietary MoBA (Mask of Bidirectional Attention) mechanism, which lets it learn world evolution in chronological order and reduce the compounding errors that cause texture blurring, geometry collapse, and scene breakdown in long-horizon generation . On the interaction side, users can drive characters with keyboard input to attack, shoot arrows, cast spells, jump, and glide, and can trigger global events like day-night cycles and weather changes through text commands. A dual-agent setup pairs a Pilot Agent that plans and executes character behaviors with a Director Agent that introduces new events as the scene progresses , and the system supports multiple users within a single persistent world for collaborative exploration . The release ships with a distilled fast inference version, an optimized generation pipeline, and day-0 support for SGLang , available on GitHub, Hugging Face, and the Reactor platform for browser trials. It arrives alongside LingBot-Video, described as an open-source video generation foundation model built on a Mixture-of-Experts architecture designed for embodied intelligence , signaling how Robbyant plans to connect generative world models to its broader work on foundational large models for embodied AI and robotic companions for use cases like elderly care, medical assistance, and household tasks .
Comments (7)
One brain for all robots is a bold pitch when my Roomba still gets stuck on the same sock every Tuesday. Show me it navigating my hallway before it summons storms.
The tweet buries the lede under three emojis and a truncated link. If the world runs for hours, lead with the hours, not the spellcasting.
Pre-recorded demo or live? Because 'runs for hours' and 'responds to every move' are two very different infra bills.
Also, 'open-source' doing heavy lifting here. Weights, inference code, or just a readme with a discord invite?
Cuts in the demo reel are way too fast to trust. When a world model is actually persistent you slow down and let one shot breathe for 20 seconds.
Every generation gets the Minecraft it deserves. Ours hallucinates.
Curious how you're framing the leap from world model to robot foundation model, those are pretty different loss landscapes. The tagline and the demo feel like two separate launches stapled together.