Ant Lingbo, Amap and Daxiao Robotics Push World Models Toward Single GPUs, 3D Cities and Robot Skills

The LingBot-World 2.0 lightweight version is designed for local real-time generation on one consumer GPU, according to QbitAI. Ant Lingbo open-sourced the 14B main model in July 2026; that model supports hour-level continuous generation, action and event interaction, and 720p/60fps real-time experience under appropriate compute and inference configurations. The 1.3B figure had appeared in the paper abstract, and AMD Senior Vice President Jack Huynh named LingBot-World 2.0 among three open-source models, alongside Zhipu and Qwen, during his opening keynote at IFA Berlin. In a demo, a prompt generated a world in which a character could explore and act for long periods in a Chinese town and European countryside. Huynh said, 'This is the future of Personal AI.'

Ant Lingbo says the smaller model was not produced by pruning the 14B model. The team first trained Causal World, a causal pretraining model that can use only past visual information and user input, then designed MoBA, a mixture of bidirectional and autoregressive attention masks, to reduce overfitting and visual degradation over long contexts. The pretrained backbone was used as a teacher, followed by consistency distillation to compress denoising steps and distribution matching distillation on the student's own long self-rollout trajectories to reduce cumulative drift. The company also released the Causal Pretrain and bidirectional Teacher models so the community can carry out post-training, distillation, compression and vertical adaptation.

Amap's ABot-Earth 0.7 uses a 3D-native technical path and spatiotemporal data to generate 3D Gaussian Splatting city scenes end-to-end. Given a satellite image or text description, it can generate a kilometer-scale 3D city scene in 10 minutes on a consumer GPU, which Amap says is 1,000 times more efficient than traditional methods. The model covers more than 196 countries and regions and supports continuous AI generation from planet to city to street-level landmarks, with free exploration and real-time interaction. Amap CEO Guo Ning said large models understand language while Amap's spatial intelligence understands the world. The capability has been applied to Flight Street View 2.0, and Amap says its spatial intelligence framework includes three layers: 3D spatial representation, dynamic perception and spatiotemporal reasoning. Amap, China's largest navigation app, reports a daily peak of nearly 1 trillion BeiDou satellite positioning calls and coverage of more than 200 countries and regions. Mercedes-Benz was named the first global automotive partner of Amap's Saostreet list.

The HSImul3R framework addresses the gap between visually accurate 3D reconstruction and physically executable interaction. It is described as the first framework for stable, simulation-ready human-scene interaction reconstruction from uncalibrated sparse views, and also supports monocular video input. Its Physics-in-the-Loop bidirectional optimization uses scene-targeted reinforcement learning to optimize human motion and Direct Simulation Reward Optimization to refine 3D scenes with simulation feedback. The team built HSIBench, which contains 19 scene objects, more than 50 human motion sequences and 300 independent interaction instances, completed by three participants and captured from 16 synchronized views per case. In Easy, Medium and Hard tasks, HSImul3R achieved interaction stability rates of 53.68%, 30.56% and 13.92%, compared with HSfM's 10.52%, 4.50% and 2.66%. The human-scene 3D clipping rate fell from 69.51% to 22.90%.

The researchers transferred optimized human motions to a Unitree G1 humanoid robot, trained whole-body control policies in simulation and deployed them to the real robot, linking image or video input to 3D reconstruction, physical simulation, robot motion transfer and real execution. The team says the route is still early, with challenges in complex interactions, multi-object scenes and dynamic environments, but it points to a way to turn large amounts of human video into simulation-ready, learnable and executable robot skill assets.