AMAP CV Lab researchers Fan Jiang and 40 coauthors published a technical report on July 21st detailing ABot-World-0, an action-controlled video model designed to generate a continuously explorable world on a single Nvidia RTX 5090 desktop GPU.
AK (@_akhaliq) highlighted the work in a thread on X on July 22nd. The report followed AMAP CV Lab's July 9th release of the model weights, inference code, local interface and online playground.
ABot-World-0 accepts keyboard commands for scene navigation and third-person character control, then generates the next frames in response. AMAP CV Lab reports output at 1280 by 704 resolution, which it describes as 720p, with 1.2 seconds between a keypress and the first corresponding decoded frame.
The system's main contribution is its deployment profile. World models have typically required costly inference infrastructure or produced short clips that offer limited user control. AMAP CV Lab built ABot-World-0 around a 5 billion-parameter causal video model and optimized the surrounding inference stack so it could fit and run on one high-end desktop card.
The 16 FPS result requires the full optimization stack
AMAP CV Lab's headline figure of up to 16 frames per second represents the upper end of several low-bit configurations, rather than the performance of its default setup.
In the researchers' system breakdown, the unoptimized model and a version using only the SageAttention2 attention implementation both ran out of memory on the RTX 5090. Adding a smaller video decoder made the pipeline run at 9.1 FPS while consuming about 20.5 GiB of peak GPU memory.
Quantizing the diffusion transformer to FP8 increased performance to 12.4 FPS and reduced peak memory use to roughly 15.9 GiB. Adding the team's Fast-RoPE positional-encoding optimization raised throughput to 13.3 FPS, while measured peak memory increased to about 19.3 GiB. More aggressive low-bit configurations produced the reported 16 FPS ceiling.
That distinction matters because AMAP CV Lab describes FP8 as its quality-oriented default. The faster result occupies a different point in the trade-off among image quality, throughput and memory use. The paper does not provide independent testing of those operating points.
ABot-World-0 also keeps memory use from growing with every generated frame by retaining a limited local context and evicting older entries from its key-value cache. This makes an open-ended session technically possible within a fixed memory budget, although the model does not preserve its complete visual history indefinitely.
LongForcing targets the drift that breaks generated worlds
Autoregressive video models can degrade as their own output becomes the input for subsequent frames. Small errors compound until objects change identity, movement freezes or a scene loses its structure.
AMAP CV Lab's LongForcing training method exposes the student model to longer self-generated rollouts while using an extended-horizon teacher model for correction. The researchers report that this reduced accumulated visual errors in a 60-second test compared with their shorter-horizon baseline. They also present selected checkpoints from hour- and day-scale rollouts that retained recognizable scenes and active motion.
Those long-run demonstrations remain qualitative and were selected by the authors. The report's quantitative evaluation uses WorldRoamBench, where AMAP CV Lab compared ABot-World-0 with systems including Google's Genie 3, Alibaba's HappyOyster, LingBot-World and HY-World 1.5. The authors report competitive scores for action following, visual quality, physical behavior and memory, but the results have not been independently reproduced.
AMAP released the model before the full data stack
The ABot-World repository and 5B model weights are available under the Apache 2.0 license. The released model is based on Wan2.2-TI2V-5B, according to its model card, and the repository includes instructions for running a local Gradio interface on Ubuntu with CUDA.
AMAP CV Lab says it trained the system with video and action data drawn from AAA games, simulation engines and internet video. Its pipeline uses automated collection, 14 deterministic quality checks, vision-language model assessment and synchronized action and text labels.
The planned 500-hour action-annotated training dataset and the bidirectional teacher model were still listed as unreleased on the project's roadmap as of July 22nd. That leaves outside researchers able to inspect and run the causal student, while the assets needed to reproduce the complete training process remain unavailable.
ABot-World-0's immediate advance is therefore practical: AMAP CV Lab has compressed an interactive video generator into a hardware envelope available to individual developers and research labs. Its usefulness beyond demonstrations will depend on whether the model can retain state, physical consistency and responsiveness across uncurated sessions, where selected long-rollout checkpoints offer only partial evidence.