NVIDIA releases Cosmos 3 Edge for on-device physical AI

NVIDIA has released Cosmos 3 Edge, a 4-billion-parameter open world model built for local physical AI. The release matters because it moves a slice of Cosmos 3 from data-center-scale simulation into hardware closer to robots, cameras, factories, and autonomous systems.

The company says Cosmos 3 Edge can run across NVIDIA Jetson, RTX PRO, DGX, GeForce RTX, and related edge systems. Its job is narrower than a general chatbot: understand a physical scene, reason about what may happen next, and generate actions or visual consequences fast enough for local robotics and vision workflows.

Cosmos 3 Edge targets real-time robot reasoning

Cosmos 3 Edge extends the broader Cosmos 3 family, which NVIDIA describes as an omnimodal world model platform for physical AI. The Edge variant is the compact deployment tier: 4B parameters, designed for memory-efficient inference, and positioned for robotics, autonomous vehicles, smart infrastructure, and industrial vision systems.

According to NVIDIA's Hugging Face launch post, the model can operate as a small vision-language model or as a post-trained world action model. In the action-model setup, NVIDIA says Cosmos 3 Edge runs at 640x360 robot-control observations, generates 32 actions per inference, and reaches 15 Hz real-time control on Jetson Thor.

That does not make it a drop-in robot brain. It gives robotics teams an open starting point for scene understanding, action-conditioned prediction, and policy post-training, especially where sending sensor streams to a cloud model is too slow, expensive, or operationally sensitive.

One architecture spans text, image, video, audio, and action

The technical hook is NVIDIA's Mixture-of-Transformers design. Cosmos 3 uses an autoregressive tower for text and reasoning tokens, plus a diffusion tower for image, video, audio, and action generation. Shared multimodal attention links those towers so the model can use a common representation of language, visuals, motion, and control.

For builders, the action format is the practical detail to check. The model card lists supported action representations for camera motion, autonomous vehicles, egocentric motion, Franka Panda arms, dual-arm setups, Agibot, UR, Google robot, WidowX 250, and UMI. If your embodiment does not map cleanly to those formats, the release is more a research base than a near-term production component.

NVIDIA also released Cosmos 3 Edge Policy (DROID), a post-trained manipulation policy for pick-and-place tasks, along with scripts intended to help developers adapt Cosmos models to their own robots, sensors, and environments.

Availability comes with hardware and validation caveats

The Cosmos 3 Edge model card on Hugging Face lists the checkpoint as public, non-gated, and ready for commercial and non-commercial use under the OpenMDW1.1 license. It also states that the model is optimized for NVIDIA GPU-accelerated systems and notes Linux as the tested operating system.

The safety caveat is straightforward: physical AI systems need use-case-specific validation before deployment. NVIDIA's own model card says integrated systems require iterative testing at unit and system levels. That is especially important for robots, vehicles, public-space cameras, and industrial automation, where a wrong action can have physical consequences.

The benchmark claims are promising but should be read as release-time evidence, not an independent guarantee. NVIDIA says Cosmos 3 Edge ranks No. 1 on VANTAGE-Bench in its parameter class for vision analytics and reaches state-of-the-art robot policy learning after post-training. Independent testing will matter most when teams try the model on their own sensor data and hardware.

What builders should check next

Cosmos 3 Edge is worth tracking if you build local robot policies, edge video agents, factory inspection systems, autonomous-vehicle research stacks, or synthetic-data pipelines for physical AI. The near-term question is not whether it replaces larger Cosmos 3 models; it is whether a compact open checkpoint can reduce the gap between world-model research and edge deployment.

Before using it, check the Hugging Face model card for supported inputs, action dimensions, license terms, hardware assumptions, and known limitations. Teams already working with agentic AI systems may also want to pair this release with LinkLoot's guide to AI agent tools, especially when deciding which components should run locally and which should stay in cloud workflows.

Evidence

NVIDIA announced Cosmos 3 Edge in its SIGGRAPH 2026 coverage and published a dedicated Hugging Face launch post. The model card confirms the public checkpoint, last-modified status, license marker, parameter count, supported input/output modes, and deployment notes. MarkTechPost and Pure AI independently covered the release and its robotics/edge positioning.