merve

NVIDIA

NVIDIA Developer

<p>NVIDIA used SIGGRAPH 2026 to push on three fronts simultaneously: agentic AI inside creative pipelines, physical AI for edge hardware, and a new open model that punches well above its weight on hardware design tasks.</p>
<h2>MCP integrations across major creative tools</h2>
<p>The most immediately practical announcement for studios is Model Context Protocol (MCP) support across Adobe, Blender, Unreal Engine, Houdini, and Affinity. The idea is straightforward: AI agents can now operate directly inside production pipelines rather than alongside them. Whether that translates to real workflow gains or just more automation debt is something studios will figure out over the next year.</p>
<p>For newsrooms, NVIDIA also released a Synthetic Video Detector NIM microservice that identifies AI-generated video with up to 92% accuracy. Given how fast synthetic media is moving, that kind of tooling is going to matter.</p>
<h2>Cosmos 3 Edge: a 4B world model for Jetson and RTX</h2>
<p>Cosmos 3 Edge is a 4-billion-parameter open world model designed to run on Jetson and RTX hardware. It watches video streams and reasons about the physical mechanics in them — useful for robotics, autonomous vehicles, and smart infrastructure where you need that kind of spatial understanding at the edge rather than in a data center.</p>
<p>The fact that it runs locally on consumer and embedded hardware is the interesting part. A world model that requires a cloud call isn’t much use on a robot.</p>
<h2>Nemotron 3 Ultra: 550B parameters, fits on a DGX Spark</h2>
<p>Nemotron 3 Ultra is a 550B-parameter Mixture-of-Experts Hybrid Mamba-Attention model, and it’s getting attention for two reasons: performance and efficiency.</p>
<p>On agentic RTL (Register Transfer Level) coding — the kind of hardware design work that feeds into chip development — it hits a 97.1% average pass rate on the CVDP benchmark across nine task categories. That’s ahead of GLM 5.2 (92.1%) and Kimi K2.6 (95.2%). More notably, it uses 71% fewer tokens than Kimi K2.6 to get there.</p>
<p>The model works with an agent framework called ACE-RTL, which runs a generate-test-reflect loop: a generator, reflector, and coordinator take turns iterating on RTL errors using simulation feedback. NVIDIA trained it on a synthetic data pipeline covering spec-to-RTL generation, code editing, and debugging, including deliberately injected hardware bugs to make the training realistic.</p>
<p>It’s available as open weights and integrates with EDA partners Cadence, Siemens, and Synopsys. The NVIDIA Agent Toolkit on DGX Station bundles it alongside NemoClaw and Omniverse libraries for local agentic workflows.</p>
<p>Soumith Chintala’s reaction — “that it fits on a DGX Spark is chef’s kiss” — captures why this matters. A 550B model that runs locally changes what’s practical for on-premises hardware design work.</p>
<h2>Research</h2>
<p>On the research side, NVIDIA presented MotionBricks for real-time character motion, ArtiFixer for 3D scene reconstruction, and new physics simulation work in the Newton engine. These are longer-horizon bets, but the Newton work in particular fits the broader physical AI story NVIDIA is building around Cosmos and Isaac.</p>


Collections

NVIDIA's SIGGRAPH 2026 announcements span three areas: agentic AI integrations via MCP into tools like Adobe, Blender, and Unreal Engine; Cosmos 3 Edge, a 4B-parameter open world model for robotics and autonomous vehicles on Jetson/RTX hardware; and a Synthetic Video Detector NIM microservice claiming 92% accuracy for identifying AI-generated video. On the research side, notable papers include MotionBricks for real-time character motion, ArtiFixer for 3D scene reconstruction, and Newton engine physics updates. A local Agent Toolkit bundling Nemotron 3 Ultra (550B parameters) running on DGX Station hardware is highlighted as a significant capability shift for on-premise AI workflows.

NVIDIA's SIGGRAPH 2026 announcements: agentic AI, Cosmos 3 Edge, and new graphics research