What is new versus the existing Agent Toolkit?

The Agent Toolkit was already NVIDIA's harness for building enterprise AI agents. The July 26, 2026 expansion adds PhysicsNeMo (AI physics models) and updated CUDA-X libraries (new sparse solvers and quantum chemistry kernels) as agent-ready skills, so agents can now directly call physics simulation and math primitives instead of bolting them on.

Which industries are the target?

Chip design and engineering. NVIDIA calls out Cadence, Siemens and Synopsys — the three major EDA vendors — plus Samsung, Silvaco, Keysight, TSMC and startup ChipAgents as launch users. The reference agent, ACE-RTL, is aimed squarely at register-transfer-level chip design and verification.

What model do the reference agents use?

NVIDIA highlights Nemotron 3 Ultra, its open-weights 550B mixture-of-experts model, as leading among open models on agentic RTL coding. ACE-RTL is the NVIDIA Research hardware-design agent built on top of it. Customers can substitute their own model — the toolkit itself is model-agnostic.

Is any of this open-source?

PhysicsNeMo, CUDA-X and Nemotron 3 Ultra are all published by NVIDIA with open weights or open source components on GitHub and Hugging Face. The Agent Toolkit itself is a commercial NVIDIA product; the newly-added libraries and the CVDP benchmark used to evaluate ACE-RTL are freely accessible.

NVIDIA · 2026-07-26 · major

NVIDIA Agent Toolkit — PhysicsNeMo and CUDA-X plug into agents for chip and physics work

NVIDIA plugged PhysicsNeMo and the CUDA-X libraries into its Agent Toolkit so autonomous AI engineers can now call physics models, sparse solvers, and quantum chemistry kernels directly, with Cadence, Siemens, Synopsys, Samsung and TSMC as launch users.

NVIDIA Agent Toolkit expansion at DAC with PhysicsNeMo and CUDA-X libraries

NVIDIA's Agent Toolkit gains PhysicsNeMo and CUDA-X libraries as agent-ready skills, aimed at automating chip design and physics workflows.

Quick facts

Maker	NVIDIA
What's new	PhysicsNeMo + CUDA-X libraries added to Agent Toolkit
New CUDA-X libraries	cuISS (iterative sparse), cuDSS (direct sparse), cuEST (quantum chemistry)
Reference model	Nemotron 3 Ultra — leads open models on agentic RTL coding
Reference agent	ACE-RTL — hardware-design agent from NVIDIA Research
Launch customers	Cadence, Siemens, Synopsys, Samsung, Silvaco, Keysight, TSMC, ChipAgents

What is it?

The Agent Toolkit expansion turns PhysicsNeMo and updated CUDA-X libraries into skills that agents can call. PhysicsNeMo brings AI physics models for design and simulation. CUDA-X adds new iterative and direct sparse solvers (cuISS, cuDSS) plus quantum chemistry tools (cuEST). Together they extend the toolkit from general enterprise agents into hard engineering domains.

How does it work?

Agent Toolkit already provided the harness for building autonomous AI engineers on NVIDIA infrastructure. The new libraries register as tools the agents can invoke, so a workflow like chip verification, structural simulation or molecular modeling can be delegated to an agent that in turn calls accelerated physics or math kernels. NVIDIA points at Nemotron 3 Ultra as its reference model and ACE-RTL, from NVIDIA Research, as a reference hardware-design agent on top of it.

Why does it matter?

Engineering software vendors — Cadence, Siemens, Synopsys — plus foundries like TSMC and system houses like Samsung are the named launch users. That signals the EDA and semiconductor stack is where NVIDIA is pushing autonomous agents next, replacing bespoke scripting around solvers with agents that can plan, call the solver, and iterate on the result.

Who is it for?

chip design and EDA teams, physics simulation groups, engineering software vendors

Frequently asked questions

What is new versus the existing Agent Toolkit?: The Agent Toolkit was already NVIDIA's harness for building enterprise AI agents. The July 26, 2026 expansion adds PhysicsNeMo (AI physics models) and updated CUDA-X libraries (new sparse solvers and quantum chemistry kernels) as agent-ready skills, so agents can now directly call physics simulation and math primitives instead of bolting them on.
Which industries are the target?: Chip design and engineering. NVIDIA calls out Cadence, Siemens and Synopsys — the three major EDA vendors — plus Samsung, Silvaco, Keysight, TSMC and startup ChipAgents as launch users. The reference agent, ACE-RTL, is aimed squarely at register-transfer-level chip design and verification.
What model do the reference agents use?: NVIDIA highlights Nemotron 3 Ultra, its open-weights 550B mixture-of-experts model, as leading among open models on agentic RTL coding. ACE-RTL is the NVIDIA Research hardware-design agent built on top of it. Customers can substitute their own model — the toolkit itself is model-agnostic.
Is any of this open-source?: PhysicsNeMo, CUDA-X and Nemotron 3 Ultra are all published by NVIDIA with open weights or open source components on GitHub and Hugging Face. The Agent Toolkit itself is a commercial NVIDIA product; the newly-added libraries and the CVDP benchmark used to evaluate ACE-RTL are freely accessible.

Try it

https://developer.nvidia.com/physicsnemo