Alibaba's Amap runs world model on a single GPU

Alibaba's location-based services platform, Amap, has released ABot-World-0, an interactive world model that runs a continuous session for 24 hours on one consumer graphics card. While typical models degrade within 30 seconds to a minute—and Amap's July 16, 2026 announcement capped performance at over an hour—this system maintains stability for a full day. The team published a seekable 24-hour record with navigation points at the six-, twelve-, and eighteen-hour marks, alongside five other runs through desert, grassland, city, and snowfield environments.

The breakthrough relies on a training method called LongForcing, detailed in a July 21, 2026 technical report. While standard distillation trains a causal student model on short clips, LongForcing applies teacher supervision to the later stages of long, student-generated rollouts. This corrects what the team describes as "accumulated distribution shift and autoregressive drift" in their report. Running on one Nvidia RTX 5090 desktop card, the optimized low-bit ABot-World-0 achieves 720p output at up to 16 frames per second, requiring roughly 19 GiB of peak video memory with a 1.2-second latency between keyboard input and frame updates.

Evaluated on the WorldRoamBench suite, the model demonstrates coherent long-horizon evolution. The project became the top paper on Hugging Face on July 22, 2026, securing over 300 upvotes. Amap has open-sourced the project under an Apache 2.0 license, releasing the inference code, a local demo, and model checkpoints on Hugging Face and ModelScope. They also shared their training dataset of 30,969 action-conditioned video episodes totaling 2.74 TB, complete with MP4 files, keyboard actions, captions, and sparse camera-pose reconstructions.

For AI practitioners, this shifts long-horizon interactive generation from expensive data-center clusters to local desktop hardware. This democratization is highly valuable for embodied AI, where developers must test control policies in simulated environments before deploying to physical machines—similar to how Google's Gemini Robotics ER 2 mimics video-action models on Audi's factory floor. By placing ABot-World-0 alongside other open-weight releases like Thinking Machines Lab's Inkling, Amap allows researchers to run, test, and build upon advanced world models without enterprise-level budgets.