Ant Group's LingBot-World 2.0 Runs Open-Source World Models for a Full Hour
Robbyant open-sources LingBot-World 2.0, a 14B interactive world model that generates coherent, playable environments for over an hour without visual degradation
Takeaways
- Open-source release: Robbyant (Ant Group's embodied AI team) releases LingBot-World 2.0, a 14B interactive world model that generates coherent video worlds for over an hour.
- Breaks the time barrier: All prior open world models degrade within minutes; LingBot-World 2.0 ran a 60-minute uninterrupted session with zero visible quality drift.
- Key architecture: MoBA attention mask + causal pretraining + two-stage distillation (consistency + distribution matching) eliminates the compounding error problem in autoregressive generation.
- Director-Pilot agents: A VLM Director proposes world events; a diffusion Pilot renders them, enabling combat, weather changes, entity spawning, and multiplayer in one persistent world.
- Real-time 720p/60fps: Distilled fast model available now on Hugging Face; try without setup on Reactor. License: CC BY-NC-SA 4.0 (non-commercial).
- Limitations: No true long-term memory (revisited regions are regenerated), occasional physics artifacts, and real-time use still requires significant GPU resources.
LingBot-World 2.0 (also called LingBot-World-Infinity) is now open-source. The release comes from Robbyant, the embodied AI research team within Ant Group, and it targets one of the most stubborn problems in generative world modeling: the clock runs out. Every prior open model degrades within seconds to a few minutes as generation errors snowball. Robbyant's new release claims to have broken that ceiling entirely, running a single uninterrupted session for over an hour with no visible quality loss.
The problem every world model hits
World models generate video frame-by-frame, conditioning each new frame on everything that came before. That feedback loop is also their weakness. Compounding errors typically cause texture blurring, geometry collapse, and scene breakdown in long-horizon generation. Most systems stay coherent for a few minutes at best, which is fine for demos but useless for anything that needs to run continuously, like a robot training environment or a persistent multiplayer game world.
While LingBot-World 1.0 set a benchmark with minutes-level stable generation, version 2.0 pushes the boundaries toward true infinity. The team stress-tested this claim by running a single uninterrupted 60-minute session across 20 distinct scenarios, with no perceptible quality decay from start to finish.
Four things that are actually new
LingBot-World 2.0 introduces four distinct upgrades: an unbounded interaction horizon from a causal pretraining paradigm, rapid response time sufficient to drive 720p video streams at 60 fps, highly diverse interactive elements including attacking, archery, spell-casting, and shooting, and an agentic harness where a pilot agent plans character behaviors while a director agent synthesizes novel environmental elements as the scene progresses.
The interaction vocabulary is a meaningful step up from prior work. Beyond navigation, users can:
- Perform combat actions: melee attacks, archery, spell-casting, shooting
- Trigger text-driven environmental events: weather shifts, day/night transitions, entity spawning
- Play in multiplayer: multiple users sharing one persistent generated world
- Control the world as a "creator" by typing arbitrary events into the Director agent