TEST RECORD · ← all field reports

FIELD REPORTS/FIELD NOTE/ISSUE #83

Steer an endless simulated world in real time

The category says video, but nothing here plays back. You steer, and the next second of the world is generated while you watch. The hosted playground makes this the rare world model you can actually try without installing anything.

MODELABot-World-0-5B-LF
PUBLISHEDAugust 3, 2026
READ TIME2 min
TESTED BYNeural Expedition
CATEGORYFIELD NOTE

Field notes

01What it does

ABot-World, from AMap's computer vision lab, turns a starting image into a place you can move through. You press direction controls and it generates the next stretch of world at 720p and 16 frames per second, with about a second of delay between your input and the picture. There is no fixed clip length: the rollout continues as long as you keep steering, and when you head somewhere the starting image never showed, the model invents coherent new scenery instead of looping back to what it already knows.

The practical surprise is where it runs. The team distilled a Wan2.2-based teacher into a small causal student that fits in 19GB of memory, so the whole live simulation works on one high-end desktop card rather than a server cluster. That moves playable world models from lab demos into territory a well-equipped enthusiast can reach.

02How to try it

The hosted playground, ABot World Studio, runs in the browser: pick a starting scene, then drive through it with the on-screen controls and see how long the world stays coherent as you explore. If you have a 5090-class GPU, the GitHub repo ships a local Gradio demo with the same interactive loop; the checkpoint downloads in one command.

03Caveat

This is the distilled student model, and "infinite" is a training claim rather than a guarantee: drift and detail loss over long rollouts are exactly what to test for. Local use assumes roughly 19GB of free video memory and was validated on an RTX 5090, so the hosted playground is the realistic path for most readers. The teacher model and the 500-hour action-annotated training dataset are on the roadmap but not released yet.

04What you can do with it

  • Explore a concept image as a navigable environment for a pitch or mood check.
  • Scout camera moves through a scene before filming or rendering anything.
  • Prototype the feel of a game level without building the level.
  • Generate endless driving or walking footage from a single frame.
  • Probe how far a world model can wander before it forgets where it started.

Try the demo

View model page

Read this issue on neuralexpedition.com →