Amap Sustains World Model Inference for 24 Hours on Nvidia RTX 5090

Interactive world models have long faced a hard physical ceiling: they degrade rapidly. Because these models generate video frame by frame, every new chunk relies directly on the visual context produced moments before. Small calculation errors compound quickly during autoregressive inference, causing scenes to collapse into noise or loop awkwardly after roughly 30 seconds to a minute of runtime.

Amap, the location-based services platform of Alibaba, pushed past that barrier by fundamentally changing how its models are supervised. Rather than relying solely on short clip distillation, the research team implemented a training mechanism called LongForcing to maintain temporal coherence and physical consistency across extended generation sessions.

How LongForcing Stabilizes 24-Hour World Generation

Traditional world models use a bidirectional teacher model to supervise a faster causal student model over brief training clips. While this teaches the student to react quickly in real time, it leaves it unequipped to handle the distribution drift that sets in during long rollouts. The LongForcing method extends that supervision directly to the later portions of autonomous generations, where accumulated errors do the most damage.

Rather than asking the model to memorize past frames through an expanded memory mechanism, the training objective pulls the model’s trajectory back toward a stable world distribution as it runs. According to technical documentation released by the platform, this correction prevents autoregressive drift and allows the model to expand into entirely new environments—such as shifting from deserts to snowfields—without requiring fresh user prompts.

The breakthrough allows ABot-World-0 to maintain image quality and interaction responsiveness across a full day of continuous operation. The engineering team published the complete 24-hour run as an inspectable, seekable record featuring fixed entry points and unedited rollouts through diverse terrains.

Hardware Requirements and Open-Source Availability

Perhaps the most significant aspect of the release is the hardware footprint required to run it. Long-horizon interactive simulation has traditionally demanded cluster-level computing resources and data-center accelerators. Amap’s optimized low-bit configurations run locally on a single consumer desktop card.

Measurements published by the platform indicate that ABot-World-0 delivers 720p output at up to 16 frames per second on a single Nvidia RTX 5090 desktop card. The setup requires roughly 19 GiB of peak video memory, with a latency of 1.2 seconds between an input action and the resulting frame. Control is handled via raw keyboard inputs for character movement and scene roaming, backed by a reference-character memory that preserves character appearance over long sessions.

The project is open-source under an Apache 2.0 license. Developers and researchers can access the inference code, local demo, and released checkpoints through the project’s GitHub repository, along with mirrors on Hugging Face and ModelScope. Additionally, the team published its training dataset, consisting of 30,969 action-conditioned video episodes totaling 2.74 TB of data complete with MP4 recordings, keyboard action logs, captions, and sparse camera-pose reconstructions.

Integrating 3D Scenes and Spatial Teleportation

Alongside the video-generation model, the platform unveiled ABot-3DWorld-0, which shares an inference pathway with the previously released ABot-Earth0.5 framework. This 3D component generates Gaussian Splatting assets across indoor, street-level, and aerial scales, complete with photorealistic textures and physical boundaries.

Amap combines its video and 3D generation models into a unified product governed by a modular generate-evaluate-repair pipeline for quality control. The system features a spatial teleport mechanism that establishes anchor points between independently generated 3D worlds. When a user traverses these points, the system unloads the current environment and loads the destination while smoothing perspective and lighting transitions.

These capabilities point toward practical use cases in embodied AI, where developers can train robotic navigation policies in simulated environments before deploying them to physical hardware. Other targeted applications include film pre-visualization and immersive educational tools, opening expandable exploration networks for independent creators operating without large enterprise budgets.