NVIDIA Releases Cosmos 3 Edge On-Device Robot Model

NVIDIA released Cosmos 3 Edge on May 31, 2026, a 4-billion-parameter open world model engineered to operate directly on devices. This model enables robots and vision AI agents to comprehend their surroundings, perform real-time reasoning, and generate robotic actions locally, addressing the need for data center-level performance in memory-constrained edge systems. Cosmos 3 Edge is the smallest tier in the Cosmos 3 family, which also includes the Cosmos 3 Nano (16B) and Cosmos 3 Super (64B) models.

A world model, as implemented in Cosmos 3 Edge, learns how an environment evolves over time by representing objects, motion, spatial relationships, and the consequences of actions. This allows systems to understand an object's location, track its movement, and predict the outcome of interactions. For instance, a robot can use a world model to predict the visual result of reaching for an object, infer the action that caused a specific change, or generate a sequence of actions to achieve a defined goal. Cosmos 3 Edge integrates these capabilities into a single, on-device model, facilitating a shared representation for understanding the current world state, simulating future possibilities, and linking these simulations to actionable commands.

The architecture of Cosmos 3 Edge, detailed in NVIDIA's technical report, features a Mixture-of-Transformers design with two distinct towers. The autoregressive tower is responsible for processing vision and text tokens, enabling comprehension and reasoning. Concurrently, the diffusion tower handles vision, audio, and action tokens, supporting prediction, generation, and neural simulation tasks. Both towers maintain separate normalization layers and multi-head attention mechanisms, contributing to their specialized functions within the model.