@aigclink: Robbyant, Ant Group's embodied AI company, open-sourced the world model LingBot-World 2.0—capable of running continuously for over an hour with no noticeable degradation in image quality, and equipped with a brain-cerebellum-like agentic harness that allows the world to unfold persistently in response to user actions. The previous world…

The embodied intelligence company Robbyant, under Ant Group, has open-sourced its world model LingBot-World 2.0 — capable of running continuously for over an hour with no noticeable degradation in visual quality, and equipped with a built-in agentic harness resembling a brain-cerebellum synergy, allowing the world to expand dynamically with user actions.

Older world models often suffer from a common issue: after running for tens of seconds to a few minutes, the image becomes blurry, geometry distorts, and the scene drifts or collapses. LingBot-World 2.0 now demonstrates “continuous operation for over an hour with no visible quality degradation” as evidence, explicitly stating this stability is “structural, not cherry-picked.”

Four major upgrades:

1. Infinite duration without quality loss – Other interactive world models (e.g., Google’s Genie 3, M-G 3.0) are limited to “minute-level” durations; it is the only one in the comparison table to achieve “hour-level / infinite” duration in the general domain, and the only one that simultaneously meets “high dynamics + semantic interaction + real-time + fully open source” (Genie 3, HappyOyster are closed-source).

Why no drift? Two key factors: causal pretraining + MoBA mixed attention, which introduces bidirectional attention into the teacher-forcing mask, supporting arbitrary-length autoregression while acting as a regularizer to suppress degradation; plus few-step distillation (consistency distillation + DMD), with corrections applied on the model’s own long-horizon rollout trajectories, specifically targeting accumulated drift over time.

2. Significantly broader interaction actions – No longer limited to “move forward / rotate camera,” it supports character actions such as combat, archery, spell casting, and shooting, and even allows environmental modifications — summoning snow or rain on demand.

3. A world that “lives on its own” – arguably the most interesting step. It adopts the agentic harness into world modeling: the VLM ‘Brain’ observes the current frame, performs causal reasoning, and proposes events; the video generator ‘Cerebellum’ translates semantic decisions into continuous frames. Meanwhile, a pilot agent handles character actions, and a director agent continuously introduces new environmental elements and events. Together they form a persistent closed loop — the world is no longer a sandbox that “pauses when you stop moving”; it evolves independently and also changes with you. The authors’ analogy is straightforward: just as strong models need a Codex-like harness to become capable engineers, world models also need a harness to become “alive.”

4. Multiple players can enter the same world simultaneously

Deployment is lightweight: the main model is 14B, with an additional 1.3B lightweight real-time variant that can run on a single GPU, delivering stable 720p/60fps output.

World models have been competing on visual quality for a long time — the next frontier may not be “how realistic the generation is,” but rather “whether the world can live on its own and live long enough.” Both weights and code are open-sourced; interested users can test it on Reactor or Lingguang.

#LingBotWorld #WorldInfinity #WorldModel #AntGroup #OpenSource