Alibaba’s Amap Runs a World Model for 24 Hours on One GPU

Alibaba's Amap platform announced that its interactive world model ABot-World-0 sustained a continuous 24-hour session on a single consumer GPU, publishing the full run as a seekable record. The model, which generates video frame by frame in response to user actions, uses a method called LongForcing to maintain stability, achieving up to 16 frames per second at 720p on an Nvidia RTX 5090 with 1.2 seconds of input latency and about 19 GiB of peak video memory. This performance far exceeds the typical 30-second to 1-minute stability of similar systems.

AI Models & Platforms https://www.unite.ai/series/artificial-intelligence/ Alibaba’s Amap Runs a World Model for 24 Hours on One GPU Add Unite.AI to your preferred sources on Google https://www.google.com/preferences/source?q=unite.ai Amap, Alibaba’s location-based services platform, says its interactive world model ABot-World-0 now sustains a single continuous session for as long as 24 hours on one consumer graphics card, and it has published the whole run as a seekable record https://amap-cvlab.github.io/abot-world-0-24h rather than a highlight reel. The page lets anyone jump to any second of the day-long rollout, with fixed entry points at the six-, twelve- and eighteen-hour marks, alongside five more complete runs through grassland, desert, city and snowfield scenes. The figure is worth something because of the ceiling it clears. An interactive world model generates video frame by frame in response to what the user does, so each new chunk is conditioned on frames the model itself produced moments earlier. Small errors feed forward, and the scene eventually degrades. Amap says https://www.prnewswire.com/news-releases/alibabas-amap-abot-world-0-sustains-24-hours-of-interactive-inference-on-a-single-consumer-gpu-302840965.html most systems in this class hold together for 30 seconds to a minute. Its own July 16, 2026 announcement https://www.prnewswire.com/news-releases/alibabas-amap-unveils-abot-world-ai-models-that-generate-interactive-video-and-3d-scenes-for-hours-on-a-single-gpu-302827320.html of the model put the number at more than an hour. How LongForcing keeps the run stable The method Amap credits is called LongForcing, described in the technical report https://arxiv.org/abs/2607.19191 the team posted on July 21, 2026. The usual way to build a model like this is distillation: a slower bidirectional teacher that can attend across a whole clip at once trains a faster causal student that only sees the past, which is what real-time interaction requires. That supervision usually covers short clips, so the student learns to look right for a few seconds and improvises after that. LongForcing extends the supervision to where the failures happen. The student generates a long rollout on its own, and a teacher with a longer temporal context supervises the later portions of it, where prediction errors have had the most time to compound. The report frames this as correcting accumulated distribution shift and autoregressive drift, not as a memory mechanism. The model is not being asked to recall earlier frames more accurately; it is being pulled back toward a stable world distribution as it goes. Amap says that shows up in behaviour: the model keeps expanding the environment with new scenes during a rollout instead of locking into the one it started in, and does it without the user feeding in fresh prompts along the way. What the reported figures cover The performance envelope comes from the team’s own measurements. Across optimized low-bit configurations, the report puts ABot-World-0 at 720p output and up to 16 frames per second on one Nvidia RTX 5090 desktop card, with 1.2 seconds between an input action and the first frame reflecting it, and roughly 19 GiB of peak video memory. Control runs on raw keyboard input for both scene roaming and third-person character movement, with a reference-character memory that holds a character’s appearance steady across a long session. For evaluation, the paper reports results on the WorldRoamBench suite along with extended interactive rollouts, describing the outcome as competitive controllability and coherent long-horizon world evolution. What the published 24-hour record adds is inspectability: the full timeline is there to be scrubbed through second by second, rather than compressed into a table. The report drew attention when it landed. Hugging Face’s paper page listed it as the top paper of the day https://huggingface.co/papers/2607.19191 for July 22, 2026, and it has since collected more than 300 upvotes there. What developers get The practical shift is the hardware. Long-horizon interactive generation has been a data-center workload, and Amap’s argument is that one desktop card now covers it, lowering the cost of entry for developers, studios and research groups working without cluster budgets. That matters most for embodied AI, where policies have to be exercised against varied environments before they meet real hardware, an area drawing steady work from Google’s Gemini Robotics ER 2 https://www.unite.ai/google-ships-gemini-robotics-er-2-with-multi-robot-teamwork/ to mimic robotics’ video-action models on Audi’s factory floor https://www.unite.ai/mimic-robotics-introduces-flux-mimic-to-bring-video-action-models-to-audis-factory-floor/ . Everything sits under an Apache 2.0 license in the project’s GitHub repository https://github.com/amap-cvlab/ABot-World , which carries the inference code, a local demo and pointers to the released checkpoint on Hugging Face and ModelScope. That puts ABot-World-0 with the broader run of open-weight releases reaching working developers this year, among them Thinking Machines Lab’s Inkling https://www.unite.ai/thinking-machines-lab-unveils-inkling-its-first-open-weights-multimodal-ai-model/ . Alongside the 24-hour record, the team released its training data https://huggingface.co/datasets/acvlab/ABot-World-Explorer-500h : 30,969 action-conditioned video episodes totalling 2.74 TB, each one packaging an MP4 with the keyboard actions that produced it, captions and a sparse camera-pose reconstruction of the scene. Publishing the corpus lets outside researchers train against the same material and test whether LongForcing holds up in their hands, and the project’s roadmap puts the bidirectional teacher model next out the door.