Robbyant, Ant Group's embodied AI lab, on Sept. 10, 2026, released the remaining LingBot-World 2.0 checkpoints, also called LingBot-World Infinity, including a 1.3B causal-fast variant aimed at single-GPU deployment. The drop matters for researchers who need open interactive world models they can try without a multi-GPU deployment stack, a gap that has left hour-scale causal simulators hard to run outside well-resourced labs.
The July 8, 2026, technical report, arXiv:2607.07534, describes four upgrades over the earlier LingBot-World line: an unbounded interaction horizon under causal pretraining, a distilled real-time variant the authors say can drive 720p streams at 60 fps, a broader action and text-event space, and a pilot-plus-director agentic harness. BYDFi reported the same-day 1.3B Lite framing on Sept. 10.
Robbyant says the system generates open-ended interactive worlds that remain visually coherent over long, hour-scale horizons with sub-second control latency. Those fps, latency, and coherence figures are company claims. MarkTechPost's July reconstruction notes that the public generate.py reference path runs at 480x832 across eight GPUs, while the headline 720p and 60 fps stream depends on a separate deployment refiner compiled for TensorRT.
Method claims center on Mixture of Bidirectional and Autoregressive attention plus distillation, including distribution matching over self-rollout trajectories, as the stated path to reduce long-horizon drift. The harness pairs a vision-language director for macroscopic events with a Diffusion Transformer pilot for low-level dynamics. Weights and code ship under CC BY-NC-SA 4.0 for non-commercial use.
Hugging Face hosts the collection, including the 1.3B causal-fast checkpoint modified around Sept. 7, 2026, alongside 14B causal-fast, causal-pretrain, and bidirectional variants. The lab page at technology.robbyant.com/lingbot-world-v2 still says model weights will be released even though GitHub and Hugging Face already list those checkpoints.
What remains open is independent measurement of real-time 720p latency, hour-scale visual coherence, and how the 1.3B Lite path compares with the 14B reference under shared protocols. Lab marketing, the open reference script, and outsider coverage should be weighed separately until those points are settled.