NVIDIA expands Cosmos platform with new tools for world model training

NVIDIA has expanded its Cosmos platform to provide developers with advanced tools for training, evaluating, and deploying world models. These models are essential for physical AI systems, as they enable machines to understand and simulate the complexities of the real world. The expansion includes new training tools, simulation checkpoints, and Cosmos Data Factory blueprints ^[1].

The Cosmos platform offers world model builders access to frontier open models, along with the necessary data, training recipes, and optimization tools to scale training data. It allows for the adaptation of open models such as GR00T, Alpamayo, QWEN, and Pi0.5 through fine-tuning and reinforcement learning ^[1]. Developers can also use distillation and quantization to optimize these models for high-performing inference ^[1].

One of the key components of this update is the Cosmos Data Factory, which provides tools for labeling, curating, augmenting, and evaluating data at scale. It currently offers five cloud-native blueprints available through GitHub, Microsoft Azure, Nebius, and CoreWeave ^[1]. These blueprints include:

• Autonomous Vehicle Dataset Generation: Curates driving data and adds synthetic examples of hard-to-capture scenarios.
• Image Attribute Augmentation: Varies object appearances to improve image search.
• Event Video Generation: Converts still images into synthetic videos of rare or unsafe events.
• Defect Image Generation: Creates images of defects in printed circuit boards, metal, and glass for inspection models.
• Video Data Augmentation: Modifies weather and lighting in existing footage while maintaining scene structure ^[1].

To accelerate these workflows, the platform incorporates agentic skills for data curation and synthetic generation. For example, Porsche is using Cosmos Data Factory through Voxel51 on Microsoft Azure to develop a shared autonomous vehicle data flywheel across Volkswagen Group brands ^[1].

In addition to data tools, the platform provides new Cosmos 3 Nano model variants. These are designed for simulation, allowing developers to test autonomous policies in repeatable, multi-view, multisensor scenarios that are difficult to capture in reality ^[1]. The update also includes NVIDIA Cosmos 3 NIM microservices, which can deliver up to 2x faster inference without any quality loss in validated configurations ^[1].

For vision AI developers, the new NVIDIA Metropolis Blueprint for video search and summarization (VSS) 3.3, built on Cosmos NIM microservices, allows teams to monitor 40% more video streams with the same GPU by turning one prompt into an efficient application ^[1]. This technology is already being used by ADNOC to detect safety hazards in real time across 65 sites ^[1].