What is ABot-World-0?
ABot-World-0 is an open world model from AMAP CV Lab, the computer-vision team of Alibaba's maps platform. Its pitch is simple: it turns a single NVIDIA RTX 5090 desktop GPU into a real-time interactive world simulator, streaming at 720p and 16 FPS with 1.2 seconds of latency and 19 GB of GPU memory. Most models of comparable quality need a rack of data-center GPUs.
A world model that needs rented compute is a demo; one that runs on a machine you own is a tool. That is the gap ABot-World-0 is trying to close.
How it works
ABot-World-0 is built on Alibaba's Wan2.2-TI2V video model and extended with an action-conditioning mechanism that ties every generated frame to a user input rather than letting the video run on freely. A single model handles both scene navigation and character control, and a reference-character input keeps a third-person character's appearance stable over long sessions. It was trained on action-annotated video from AAA games, simulation engines and the internet.
To stop errors from compounding as the model feeds on its own output, the team developed LongForcing, a training method that aligns the model's long self-generated rollouts with an extended-horizon teacher. The result is open-ended, “infinite” rollout: the world keeps expanding with new scenes instead of locking into the one it started with.
What is open, and what is not
Under Apache 2.0, AMAP released the distilled real-time model (ABot-World-0-5B-LF), inference code, a local Gradio demo, an online playground called ABot World Studio and a 500-hour training dataset with action annotations. The bidirectional teacher model is still on the roadmap, and training code has not been released, so fine-tuning the distilled model properly takes real engineering work.
How to try it
The quickest route is ABot World Studio in the browser. To run it locally you need a recent NVIDIA GPU — the team tested on an RTX 5090 with Ubuntu 22.04 — then clone the repository, install the requirements and download the checkpoint from Hugging Face or ModelScope.
Strengths and limitations
Strengths
- Real time on a single desktop GPU
- Apache 2.0, commercial use permitted
- Navigation and character control in one model
- Open-ended rollout that keeps generating new scenes
- Annotated training dataset released
Limitations
- Game-style actions rather than free-form semantic commands
- Open-ended imagination can wander away from a fixed plan
- Teacher model and training code not yet released
- 16 FPS is below typical game frame rates
Who is ABot-World-0 for?
Developers, indie studios and researchers who want a real-time world model running on hardware they own, and a commercially usable base to build on.
Frequently asked questions
Can ABot-World-0 run on a consumer GPU?
Yes. Its developers report 720p at 16 FPS with 1.2 seconds of latency and 19 GB of GPU memory on a single NVIDIA RTX 5090.
Is ABot-World-0 open source?
Yes, under the Apache 2.0 license. The distilled model, inference code, demos and a 500-hour annotated dataset are public; the teacher model and training code are not yet released.
What is ABot-World-0 based on?
It is built on Alibaba's Wan2.2-TI2V video model, extended with action conditioning and the LongForcing training method.
Can I fine-tune ABot-World-0?
The weights and license allow it, but there is no official training code yet, so you would need to build your own training pipeline.
Where can I try ABot-World-0 without installing it?
In ABot World Studio, the online playground run by AMAP.