NVIDIA posted its SIGGRAPH 2026 announcements on July 20, 2026, led by Cosmos 3 Edge, a compact world model for edge GPUs, and new Omniverse libraries that let AI agents build simulation-ready 3D worlds. SIGGRAPH 2026 ran July 19–23 at the Los Angeles Convention Center, according to NVIDIA’s event page.

On this page

Cosmos 3 Edge

In its SIGGRAPH roundup, NVIDIA describes Cosmos 3 Edge as “now openly available,” a 4-billion-parameter “omnimodel” optimized “for memory-efficient deployment and high throughput on NVIDIA Jetson, NVIDIA RTX PRO and NVIDIA DGX systems, as well as GeForce RTX GPUs.”

Extending Cosmos 3, the model “can understand and generate text, image, video, ambient sound and action,” NVIDIA says, and uses a mixture-of-transformers architecture for real-time vision analytics and robot action on device. NVIDIA reports that Cosmos 3 Edge ranks “No. 1 on VANTAGE-Bench for vision analytics success in its parameter class.” The model is available on Hugging Face, with inference and post-training frameworks and recipes on GitHub. NVIDIA lists Agile Robots, Doosan Robotics, Siemens, Skild AI, Centific, Vaidio and YUAN among partners evaluating it.

Agent Toolkit gets Omniverse libraries

A press release the same day added Omniverse libraries to the NVIDIA Agent Toolkit so agents can produce “simulation-ready” worlds. NVIDIA names three components:

Component What NVIDIA says it does
ovrtx Generates camera, lidar, radar and other sensor outputs from 3D scenes
ovphysx Adds GPU-accelerated physics to 3D scenes
CAD-to-SimReady skills Convert CAD data into simulation-capable assets

The libraries are openly available on GitHub, alongside a blueprint for using them inside Blender. Named adopters include SideFX, PTC’s Onshape, and the startups Palatial, Lightwheel, Moonlake AI and ForgeCAD.

MCP in creative apps

NVIDIA also highlighted creative tools adding Model Context Protocol (MCP) connections so agents can act inside them. The list in the blog includes Adobe Firefly, Express and Creative Cloud, Affinity by Canva, Blender (through a lightweight MCP server in Blender Lab), Boris FX Silhouette, Foundry Griptape, SideFX Houdini 22 and Unreal Engine.

Synthetic video detection for newsrooms

NVIDIA announced the Synthetic Video Detector NIM microservice, part of NVIDIA AI for Media. It analyzes video “frame by frame to produce a classifier score of whether it contains synthetic content.” In NVIDIA’s testing, accuracy reached up to 92% on uncompressed video, 87% at 15% compression and 82% at 50% compression; NVIDIA says it can process 1080p video in as little as 22 milliseconds on RTX systems. Wowza is embedding it in the Wowza Video Intelligence Framework.

NemoClaw on DGX Station and themed days

NVIDIA also presented NemoClaw on DGX Station as a local agent stack for designers and engineers; the desktop system is built around the GB300 Grace Blackwell Ultra Desktop Superchip (see the AI chips catalog). The event page lists themed days, including Educators’ Day (July 20), Physical AI Day (July 21), RTX Rendering Day (July 22) and AI in Production Day (July 23).

An NVIDIA staff summary on the NVIDIA Developer Forums collects the same announcements with links. Model launches are tracked in our AI model release timeline.

Sources