NVIDIA releases Cosmos 3 toolkit

NVIDIA used COMPUTEX announcements on June 1 to show how it wants developers to build AI systems that do more than answer prompts. The company published Cosmos 3 for what it calls physical AI — software that has to interpret scenes, predict changes and choose actions in robots, autonomous vehicles and smart-space systems. In parallel, NVIDIA said Jetson devices are now “agentic-ready” with JetPack 7.2 and NemoClaw support, and said OpenShell can run secure local agents on Windows PCs with RTX chips and on DGX Spark systems. The releases extend NVIDIA’s software stack from model training and simulation into systems that have to work with sensors, devices and local inference. ### What did NVIDIA actually release? NVIDIA’s developer blog said Cosmos 3 combines physical-AI reasoning, world generation and action prediction in one foundation stack. The company described it as an open release aimed at developers building systems that need to understand environments, anticipate what happens next and generate actions for specific settings such as warehouses, roads or buildings. (forums.developer.nvidia.com) The June 1 post said Cosmos 3 includes model checkpoints, code and post-training recipes, along with datasets tied to robotics, autonomous driving, physics simulation, human motion and warehouse operations. NVIDIA also said the release builds on earlier Cosmos components such as Transfer, Predict and Reason models introduced over the past year. (forums.developer.nvidia.com) ### Why does “physical AI” matter in this announcement? NVIDIA said physical-AI systems need more than language understanding because they must model motion, objects and cause-and-effect in the real world. In the company’s framing, that includes robots that move through workspaces, vehicles that anticipate traffic behavior and smart-space systems that respond to changing physical conditions. (forums.developer.nvidia.com) The March and June technical posts show NVIDIA tying Cosmos to synthetic-data generation and physics-aware world models. That suggests the company is positioning Cosmos not as a chatbot layer, but as infrastructure for developers training and adapting embodied systems. That is an inference from NVIDIA’s product descriptions and release materials. (forums.developer.nvidia.com) ### What changed on Jetson? NVIDIA said on June 1 that Jetson is now “agentic-ready” with JetPack 7.2 and NemoClaw support. The company said the combination gives developers a production-grade stack for robotics, inspection and industrial automation, and said the announcement was made at COMPUTEX. Jetson has already been part of NVIDIA’s edge-computing push, and earlier company posts said developers were using Jetson hardware to run models such as Nemotron, Cosmos and Isaac GR00T locally. (developer.nvidia.com) The new JetPack 7.2 and NemoClaw support adds a more explicit software path for agents operating on embedded systems rather than only in data centers. ### What is OpenShell doing on PCs? (blogs.nvidia.com) NVIDIA’s RTX AI Garage post said OpenShell brings secure agents to Windows on RTX PCs and DGX Spark, with 2x inference performance on llama.cpp. The company framed that as a way to run local agents privately on-device, rather than depending entirely on cloud services. Earlier NVIDIA materials described OpenShell and NemoClaw as early-preview software for policy-based privacy and security controls around autonomous agents. (blogs.nvidia.com) The June update ties that work to consumer and workstation hardware, including GeForce RTX PCs, RTX PRO systems and DGX Spark. ### How do these pieces fit together? NVIDIA’s June 1 releases line up as a cloud-to-edge stack: Cosmos 3 for world and action models, Jetson software for embedded deployment, and OpenShell for local agents on PCs. (blogs.nvidia.com) NVIDIA did not present that as a single packaged product, but the separate announcements point to one developer path across training, adaptation and on-device execution. That is an inference based on the three releases published the same day. (blogs.nvidia.com) COMPUTEX runs June 2 to June 5 in Taipei, and NVIDIA’s posts direct developers to documentation, model downloads and related software pages on its developer and company blogs. (blogs.nvidia.com) (forums.developer.nvidia.com)