Alibaba's Amap: ABot-World-0 Sustains 24 Hours of Interactive Inference on a Single Consumer GPU

BEIJING, Aug. 3, 2026 /PRNewswire/ -- Amap, Alibaba's location-based services platform, said it has extended the inference length of its interactive world model ABot-World-0 to as long as 24 hours on a single consumer-grade GPU, up from the roughly one-minute ceiling of mainstream interactive world models.

The 24-hour result further validates the long-horizon capability enabled by LongForcing, the training method introduced with ABot-World-0. LongForcing keeps image quality, physical consistency and interaction responsiveness stable across day-long generation. ABot-World-0 has been open source since its initial release and is used for interactive video, game design and simulation training, Amap said.

The related research paper topped the daily and weekly rankings on Hugging Face's paper leaderboard shortly after release.

Most interactive world models remain stable for only 30 seconds to one minute. Since autoregressive video generation conditions each new chunk on previous visual context, small errors can accumulate across repeated generation steps.

LongForcing takes a different approach. It folds long-horizon stability and open-ended scene generation into a single learning objective, and rather than asking the model to memorize earlier frames, it uses a teacher model with longer temporal context to extend the training target from short-clip quality to the distribution the model produces after many autoregressive steps.

During training, the model's outputs are fed back as inputs while being kept within a stable world distribution, reducing error accumulation and drift so that scene structure and dynamics remain coherent across day-long inference.

Under the initial multimodal conditions, user actions can drive continued scene evolution without repeated prompt switching.

With the new training, ABot-World-0 sustains continuous, physically consistent generation for 24 hours on a single consumer-grade GPU.

The capability repositions long-horizon interactive video generation from a data-center workload to one that runs on consumer hardware, lowering deployment cost for independent developers, studios and research institutions, Amap said.

The updated ABot-World-0 remains open-source and is available on Hugging Face and Reactor.

References:

ABot-World Studio: https://abot-world.amap.com/#world

GitHub: https://github.com/amap-cvlab/ABot-World

ABot-World-0 24-hour inference demo: https://amap-cvlab.github.io/abot-world-0-24h

SOURCE Amap