LingBot-World 2.0：60分钟无衰减的开源交互式世界模型

Most world models start breaking after seconds or minutes;

@robbyant_brain's newly released LingBot-World 2.0 reports a single 60-minute rollout spanning 20 distinct scenarios, with no perceptible decay from start to finish.

Also called LingBot-World-Infinity, it’s an open-source interactive video world model that takes 1 starting frame plus live camera movements or text instructions, then keeps generating the scene as the user moves and acts inside it.

A generative world model is like a base LLM — it doesn't “play itself.” LingBot-World-Infinity wraps it in a harness.

Same idea as coding agents: a strong model only becomes useful inside a scaffold that lets it inspect state, act, and chase a goal across turns. They apply that to world modeling.

Causal world models build each new frame from earlier generated frames, so 1 small mistake can get reused until textures smear, geometry bends, or the whole scene drifts;

• A frame predictor doesn't play itself. The world model needed a harness the same way a code model needs one. - Two agents run the world: a pilot for the character, a director so the scene never runs dry.

LingBot-World 2.0 targets that failure directly, with a single 60-minute run across 20 scenarios showing no perceptible quality decay, alongside real-time output at 720p and 60fps.

Training has 2 stages: a slower causal diffusion model first learns high-quality world prediction, then consistency distillation compresses it into a few-step student, while Distribution Matching Distillation trains that student on its own long rollouts so it learns from the imperfect states it will actually create during use.

A Vision-Language Model acts as the “Brain” that proposes events, the video generator acts as the “Cerebellum” that renders them, and pilot plus director agents handle character behavior and keep adding fresh events; the release also includes a 14B main model, a 1.3B model for a single consumer GPU, open weights, the agentic harness, multiplayer controls, and an online demo, though it still lacks true long-term memory and may regenerate a place instead of recalling the exact same place later.

🧵 1.