InternW0

The first model in Shanghai AI Laboratory's InternW series of physical world models, from the lab's Physical Intelligence Team (the only affiliation printed on the report). InternW0 is a world action model: a mixture-of-transformers backbone pairs a high-capacity video expert, which predicts how the scene will evolve, with a lightweight action expert that generates continuous robot control, and the two are trained jointly with video–action flow matching. Its central idea is asynchronous duplex inference. The video expert refreshes its prediction on a slow schedule and exposes its layerwise K/V as a cached plan; between refreshes, an observation-conditioned chunk K/V editor adapts that cache to each new observation, so action updates do not wait for video generation. Frames are encoded by a frozen Wan VAE and a frozen DINOv3 encoder, domain-specific interfaces and soft prompts map different robots onto a unified 37-dimensional action space, and contact-aware post-training adds force and tactile input with joint prediction of six-dimensional interaction wrenches. Pretraining uses about 7,200 hours of robot and egocentric data from seven datasets, including EgoLab, 275 hours of first-person video the team recorded in real laboratories. The report gives no parameter count and does not say how the video expert is initialized.

After task-specific post-training, InternW0 averages 98.6% on LIBERO, ahead of LingBot-VA (98.5%) and π_0.5 (96.9%). On RoboTwin 2.0 it scores 93.12% in the Full setting, 0.98 points behind ABot-M0.5, and 75.60% in the Clean2Random setting (post-trained on clean scenes, tested on randomized ones), 8.3 points above the best baseline. On five real-robot tasks it matches or beats π_0.5 on every one, including a 15-stage metal–organic framework synthesis workflow (68.4% average progress against 50.2%) and quantitative pipetting with a 20-degree-of-freedom dexterous hand under hybrid force–position control (65.3% against 46.7%). The action path takes 60.73 ms on an RTX 5090D, a 16.47 Hz update rate and 3.13× faster than Fast-WAM. Within the lab's InkStone scientific-discovery platform, InternW0 connects the reasoning and tool use of Intern-S2-Preview to physical experiments. Pandaily reported the launch on September 14, and the report was posted on September 23 (the date used here). InternW0's own weights are not public: the README of its project-page repository lists the GitHub, Hugging Face and ModelScope model links as still to come.

InternW0-Δ, a companion world action model from the same team (report posted September 25), is the part with public weights. Its video expert is initialized from Alibaba's Wan2.2-TI2V-5B, a frozen RynnBrain1.1-2B vision-language model gives a randomly initialized action expert scene-level semantics, and a Track4World teacher distills 4D geometry and motion priors into the video expert during training only. Causal Imprint queries learn future-relevant scene changes from future frames that serve only as training targets, so the action expert gets predictive features without rolling out video at inference. It is pretrained on what the team calls the largest open-source corpus of its kind, over 20,000 processed hours: 11,302 hours of robot demonstrations, 2,075 hours of UMI data, 4,061 hours of egocentric human video, and 5,634 robot-hours converted from that video (Ego2Robot), for 245K steps (about 14 days) on 256 A800 GPUs. Post-trained only on unperturbed demonstrations, it reports the best results in its comparisons: 92.8% on LIBERO-Plus, 71.9% on RoboTwin 2.0 Clean2Random and 90.0% on Clean2Clean, 66.0 on EBench, and 23.9% on RoboDojo, nearly double the strongest prior world action model. The pretrained checkpoint and post-trained LIBERO, RoboTwin and RoboDojo checkpoints are on Hugging Face under Apache 2.0, and the code, which covers data processing, training, evaluation and real-robot deployment, is MIT-licensed. Because its backbone starts from Wan2.2, InternW0-Δ is a derivative model; no parameter count is given.