Introducing Cosmos 3 Edge

Source: HuggingFace Date: 2026-07-20 URL: https://huggingface.co/blog/nvidia/cosmos3edge

Summary

Cosmos 3 Edge is NVIDIA’s 4B-parameter open world model for physical AI, purpose-built to run on robots and vision-AI agents rather than serve them from the cloud: it reasons about surroundings and generates robot actions directly on edge hardware, hitting 32 actions per inference at 15Hz real-time control on a Jetson Thor. Architecturally it pairs an autoregressive tower (understanding/reasoning) with a diffusion tower (prediction/action generation) sharing multimodal attention, and ranks #1 among 4B-parameter models on VANTAGE-Bench.

Implications

Extends NVIDIA’s open physical-AI line — Cosmos 3 opened June 1 as the “fourth open contributor” alongside Google/Alibaba/Zhipu, notable because NVIDIA’s incentive to keep it open is structural (open physical-AI models drive demand for the Blackwell/Hopper/Jetson silicon NVIDIA sells). Cosmos 3 Edge pushes that line down to edge/robot-control scale, feeding the physical-AI/robotics thread on a different axis than the chat/coding open frontier: instead of racing total-parameter scale (K3 at 2.8T), this model optimizes for tiny, real-time, on-device footprint. Worth reading alongside today’s Grabette signal — cheap open edge inference (this) plus cheap open data collection (Grabette) is the same “physical AI decouples from big-lab-only” move the chat/coding frontier made roughly a year earlier.