What is LingBot-World 2.0 (Infinity)?
LingBot-World 2.0, also called LingBot-World-Infinity, is an interactive video world model from Robbyant, Ant Group's embodied-AI company. It was released on 9 July 2026, with the remaining model variants following on 10 September. It succeeds LingBot-World from January 2026.
The main model has 14 billion parameters and is built on Alibaba's Wan2.2; a 1.3B fast version is meant for a single GPU. You control it with camera moves, actions such as attacking, archery or spell-casting, and text events. Two built-in agents can run the scene: a pilot agent plans what the character does, and a director agent adds new elements as the world unfolds. Several people can share one world.
Robbyant claims 720p at 60 frames per second and generation that runs for an hour. The licence is CC BY-NC-SA 4.0, non-commercial only, replacing the Apache 2.0 licence of the first version.
Is LingBot-World 2.0 (Infinity) worth it?
LingBot-World 2.0 is the most feature-rich open world model: hour-long sessions, combat and magic actions, text-triggered events, multiplayer and two AI agents that can play and direct the scene for you. For anyone studying how a generated game might run itself, nothing else open comes close in scope.
The first version found an audience. Its GitHub repository has 4,571 stars and the fast checkpoint had 62,637 downloads on Hugging Face in the last month; the new repository has 1,904 stars.
The catch: the headline numbers are not backed by measurements. The paper gives no GPU type or count for 720p at 60 frames per second, and that rate includes a separate refiner that upscales and adds in-between frames across several GPUs. There is no latency figure, no benchmark and no user study; the hour-long claim rests on one 60-minute session judged by eye. Robbyant's own limitations section says places you return to are regenerated rather than remembered, characters and style drift over long sessions, objects can pass through each other, and it needs substantial compute. The reference scripts use 8 GPUs for the 14B model and 4 for the 1.3B model at 480p, and Robbyant will not release its deployment code. The licence also went from Apache 2.0 to non-commercial.
Our take: read the paper for the agent design and play the Reactor demo for the feel; for anything you might sell, Matrix-Game 4.0 or Waypoint 1.5 have the licence and the numbers this one lacks.
Where it earns its place
An AI researcher studying agents in simulated worlds
Test how a planning agent and a scene-director agent interact in a generated environment
Run the 14B causal-fast model on an 8-GPU node with the released inference code, or deploy through SGLang, and log the pilot and director agents' decisions; non-commercial research only.
A game designer evaluating world models
Judge whether hour-long generated sessions are usable
Play the hosted version on Reactor and note where drift or regenerated locations break the session; no hardware needed and no code.
Before you commit
- Setup
- Clone Robbyant/lingbot-world-v2 (built on Wan2.2), download a checkpoint from Hugging Face or ModelScope, and run generate.py with torchrun; the fast 14B model uses 8 GPUs and the 1.3B model 4 GPUs in the reference scripts.
- Integrations
- SGLang and NVIDIA flashdreams for deployment; Reactor (web) and LingGuang (mobile) host demos.
- Data and privacy
- Self-hosted, nothing leaves your servers. CC BY-NC-SA 4.0: no commercial use, and derivatives must use the same licence.
- Limits
- No published latency or hardware for 60 fps; regenerates revisited places; drift and objects passing through each other; deployment code not released.
What’s good
- Pilot and director agents can play and extend the scene
- Wide action set: combat, archery, spell-casting, shooting
- Multiplayer worlds
- 1.3B variant aimed at a single GPU
What’s not
- Non-commercial CC BY-NC-SA 4.0 licence
- No hardware, latency or quality measurements published
- No real long-term memory; drift and weak physics
- Deployment code withheld; reference scripts use 4 to 8 GPUs
Best for
Researchers interested in agent-driven world models who can work under a non-commercial licence and have a multi-GPU server.
LingBot-World 2.0 (Infinity) alternatives
LingBot-World 2.0 (Infinity): common questions
- Is LingBot-World 2.0 (Infinity) worth it?
- LingBot-World 2.0 has the most ambitious design of any open world model, and the least evidence that it works as described. We rate it 3.6 out of 5.
- How much does LingBot-World 2.0 (Infinity) cost?
- Free weights for non-commercial use (CC BY-NC-SA 4.0). You pay for the GPUs.
- What are the downsides of LingBot-World 2.0 (Infinity)?
- Non-commercial CC BY-NC-SA 4.0 licence; no hardware, latency or quality measurements published; no real long-term memory; drift and weak physics; deployment code withheld; reference scripts use 4 to 8 GPUs.
- Who is LingBot-World 2.0 (Infinity) best for?
- Researchers interested in agent-driven world models who can work under a non-commercial licence and have a multi-GPU server.
- Is LingBot-World 2.0 (Infinity) safe for company data?
- Self-hosted, nothing leaves your servers. CC BY-NC-SA 4.0: no commercial use, and derivatives must use the same licence.
- What are the best LingBot-World 2.0 (Infinity) alternatives?
- The closest alternatives we have reviewed are Matrix-Game 4.0, Waypoint 1.5, HY-World 1.5 (WorldPlay) and Genie 4.
Answers checked Nov 7, 2026.