Ant Lingbo open-sources a world model it says stays coherent for an hour

Ant Lingbo, an AI lab under Ant Group, says its newly open-sourced LingBot-World 2.0 can keep a simulated, interactive scene running coherently for a full hour. That is the headline number from a stress test the company itself ran. It is also why the lab is releasing the weights, code, and a technical report rather than a polished demo: Ant Lingbo wants outside developers to test whether the claim holds.

A "world model," in the sense Ant Lingbo uses the term, is an AI system that simulates an interactive 3D scene frame by frame in real time, instead of producing a pre-recorded video clip. Think of it as a game engine running inside a neural network: the user presses a key, and the world responds. Most such systems decay within minutes. The lab's predecessor, LingBot-World 1.0, open-sourced in January 2026, held up for nearly ten minutes. Version 2.0, according to Ant Lingbo, pushes that to "hour-level."

The mechanism the company is betting on is what it calls an Agent-native design. Two agents share control of the simulation. A "Pilot Agent" drives the character the user is playing, taking keyboard or other input and steering movement, attacks, archery, magic, jumping, gliding. A "Director Agent" triggers scene-level events: day/night switches, weather changes, new entities appearing. Both run inside a single persistent world. That is what Ant Lingbo means by "AI-native multiplayer": multiple users, or multiple agents, can drop into the same scene and continue where the last session left off.

Underneath the agents sits a training paradigm Ant Lingbo calls causal pre-training, paired with a Mixture-of-Block-Attention (MoBA) mechanism developed in-house. The model is trained to predict world evolution forward in real time, second by second, instead of regenerating from scratch. That ordering is supposed to reduce compounding error during long rollouts, the usual reason video-style world models collapse after a few minutes.

The one-hour stress test was the company's own. Ant Lingbo reports that the output remained clear and structurally coherent with no obvious quality decay across a full hour of continuous interaction. There is no independent benchmark replication yet. The arXiv paper is the closest public artifact for outside scrutiny. Treat the hour figure as a vendor claim until a third-party test reproduces it.

The v2 GitHub repository holds the inference code. A fast-variant checkpoint lives on Hugging Face under the robbyant namespace, described as a 14-billion-parameter causal-fast version distilled from the pre-trained model. Its inference pipeline is tuned for stream-then-transmit-then-display, so users can act on a scene without waiting for a batch to render. The lab's Reactor platform and Lingguang (灵光) consumer app are the channels Ant Lingbo is steering end users toward.

Use cases the company names are familiar for this category: game content generation, film and TV previsualization, virtual simulation, digital twins, and training data for robotics and embodied AI. LingBot-World 2.0 sits inside Ant Lingbo's broader "full-stack brain 2.0" plan, where it is positioned as one building block rather than a finished product.

The gap to watch is the one between vendor claims and reproducible performance. Ant Lingbo's release is also a positioning move. The lab is one of several Chinese groups pushing world models this year, and "industry first" language in the announcement is the vendor's framing, not an audited benchmark result. Independent evaluation of hour-scale coherence, the quality of the Agent coordination under load, and the actual license terms on the released weights will determine whether LingBot-World 2.0 is a real step forward or a polished demo with a generous timer.

The repositories and weights went public on July 9, 2026. Outside developers are the first real test.