FULL STORY

NVIDIA Cosmos 3: From Edge Teaser to World Model Launch

NVIDIA progressively unveiled the Cosmos 3 series around SIGGRAPH, moving from the Edge visual model for robotics to the official launch of the Cosmos 3 world foundation model.

2026-07-16 ~ 2026-07-30 · 6 episodes · 25 posts

Episode 1 · NVIDIA Unveils Cosmos 3 Edge On-Device Vision Model (2026-07-16, 2 posts)

NVIDIA announced the expansion of the Cosmos Coalition to Japan and introduced Cosmos 3 Edge. The new 4-billion-parameter model is specifically designed for on-device visual inference to enhance edge AI capabilities.

Episode 2 · NVIDIA Previews SIGGRAPH 2026 and Launches Cosmos 3 Edge (2026-07-20, 15 posts)

At SIGGRAPH, NVIDIA previewed its major updates for 2026, focusing on advancing graphics and simulation by deeply integrating agentic AI with physical AI. According to @nordicinst, this move aims to build smarter graphics systems and more realistic simulation environments. This strategy covers breakthroughs in open-source models and real-time simulation, profoundly impacting future media production workflows.

Launch of Open-Source Edge World Model Cosmos 3 Edge

NVIDIA released Cosmos 3 Edge, a 4-billion-parameter open-source world model capable of running on edge GPUs. The model consists of a 4B main body and a 2B Nemotron reasoner. Designed for physical AI scenarios, it runs on local devices like Jetson Thor and can understand and generate text, images, video, environmental audio, and motion. The model is fully open-source, including model weights, post-training recipes, and code, and is available on Hugging Face. Officially, it ranks first on the VANTAGE-Bench visual analysis leaderboard among open models in its size class (under 4B parameters), surpassing previous Cosmos models. Its application scenarios include robot learning and action, autonomous driving scene understanding and intent prediction, and real-time video reasoning for visual AI agents.

Pursuing Behaviorally Authentic Physical World Generation

NVIDIA stated that its research goal is no longer limited to generating visually realistic worlds, but rather creating 3D worlds that are behaviorally authentic and respond in real-time. This AI-generated content must be 3D-based, follow the laws of physics, and allow for direct creator control. This technological direction will simultaneously serve the simulation needs of gaming, film, robotics, and factory digital twins.

Creative Software Accelerating Towards "Agent-Ready"

Beyond underlying graphics and simulation upgrades, NVIDIA highlighted a key application-layer trend: pushing leading creative applications to open Model Context Protocol (MCP) connections. This initiative allows AI agents to work directly inside creative tools across scenes, shots, timelines, assets, and editing. NVIDIA describes this shift as the next phase of "accelerated creation," signaling that creative software is evolving to become fully "agent-ready." In this new workflow, AI handles the heavy lifting deep within the tools, while human creators retain ultimate control over the creative process.

Introducing Synthetic Video Detection and Desktop Agent Stack

As AI-generated content becomes more prevalent, NVIDIA introduced the Synthetic Video Detector NIM microservice based on its AI for Media platform. Targeting media organizations like broadcasters and publishers, this service aims to provide trusted verification tools that keep pace with modern content production speeds to combat synthetic videos. Furthermore, NVIDIA announced that desktops can now run "super agents," releasing the NemoClaw open agent stack based on DGX Station, integrating technologies like Nemotron 3 Ultra and Omniverse.

Episode 3 · NVIDIA Integrates AI Agents into 3D Creation via RTX PRO and DGX (2026-07-21, 2 posts)

NVIDIA is leveraging RTX PRO and DGX systems to integrate AI agents into 3D creation workflows. By connecting these agents to Blender and Unreal, users can automate repetitive tasks and infer 3D scenes while retaining full creative control.

Episode 4 · NVIDIA Unveils Cosmos 3 Edge for On-Device Robotics on Jetson (2026-07-21, 2 posts)

NVIDIA introduced Cosmos 3 Edge, a physical AI inference model designed for edge devices like Jetson. It enables robots to analyze video streams in real-time, understand physical mechanics, and predict next actions on-device, enhancing rapid on-device robotic control.

Episode 5 · NVIDIA Releases SIGGRAPH 2026 Keynote Replay (2026-07-22, 2 posts)

NVIDIA has released the replay of its SIGGRAPH 2026 keynote. The presentation primarily focuses on the latest advancements in neural rendering, world models, and robotics simulation.

Episode 6 · NVIDIA Unveils Cosmos 3 World Foundation Models (2026-07-28, 2 posts)

NVIDIA launched the Cosmos 3 world foundation models at SIGGRAPH 2026 to enhance physical AI understanding, alongside a compact 4-billion-parameter Cosmos 3 Edge model designed for robotics.