Robbyant drops Lingbot-World 2.0 for real-time playable video

LingBot-World 2.0, open-sourced by Ant Group's Robbyant, is an interactive world model that turns generative video into playable virtual environments in real time. Instead of pre-rendering a 3D scene, it processes keyboard inputs and text prompts on the fly to stream reactive video frames.

What it can do

Most interactive video models hit a wall after a minute or two: pixels smear, geometry warps, and the scene collapses into visual noise. LingBot-World 2.0 tackles this drift head-on, running for over an hour of continuous interaction without noticeable visual breakdown.

The interaction range goes well beyond simple camera panning. It supports actions like attacking, archery, casting spells, and shooting. It can handle unexpected entities too, such as controlling a movable chair sliding through a room. You can also inject environmental events via text prompts, like summoning rain or snow mid-session.

Under the hood, it pairs the video generator with an agentic harness. A pilot agent handles character planning and execution, while a director agent introduces novel environmental elements as the scene moves forward, creating an active loop that keeps the world evolving.

Specs and how to run it

Built on top of Wan2.2, the framework uses causal pretraining combined with KV caching to process frames chunk-by-chunk. A distilled real-time variant is capable of pushing a 720p video stream at 60 fps.

Robbyant released both 14B and 1.3B parameter models. The 14B variant focuses on fidelity, while the 1.3B version targets lower latency. Weights are hosted on Hugging Face and ModelScope, totaling around 86GB across the packages.

Local hardware requirements are steep. The included inference scripts run at 480p and require 8 GPUs for the 14B causal-fast model, or 4 GPUs for the 1.3B model. The repo is licensed under CC BY-NC-SA 4.0 for non-commercial research. Robbyant has noted it will not release its internal production deployment code, pointing self-hosters toward community implementations in SGLang or flashdreams. If you prefer not to spin up a GPU cluster, hosted real-time demos are accessible via Reactor and LingGuang.

Why it matters

If you work in robotics simulation or generative entertainment, LingBot-World 2.0 is one of the rare open alternatives in a field dominated by closed demos like Google's Genie 3. Most previous open models were confined to narrow domains or collapsed within seconds.

We think this release moves world modeling past the stage of fragile video tech demos into functional, long-horizon interactive sandboxes. The multi-GPU requirement and non-commercial license will keep it out of commercial game pipelines for now, but it offers researchers an exceptional open workbench for training embodied agents and probing real-time generative physics.