NVIDIA PhysicsNeMo

Is this your product? Claim it free

Hidden from Google until claimed. Claimed listings appear in Google, carry your own description, and include 100 free custom questionnaires. The first 10 products also get a founding ranking slot with 100 review codes.

Start an assessment

TrustMark reading

Where this product sits

NVIDIA PhysicsNeMo is classified as Simulation. Software that models queues, throughput and markets so operators can try a change without touching production. Jobs: queueing and throughput trials; synthetic counterparties in markets; what-if on operational flow.

Buyers typically come to Simulation because teams cannot rehearse a process without touching production. Classification is not a finding that this product solves that job.

This is TrustMark's classification reading for buyers comparing Simulation products. It is not an Assurance Score, certificate, or ranked position.

What it is

Product classification

AI Platform / Infrastructure

NVIDIA PhysicsNeMo is an open-source Python framework for building, training and fine-tuning physics AI models at scale, combining physics knowledge with deep learning methods such as neural operators and graph networks. Engineering and research teams use it to build fast surrogate models for simulations in areas like fluid dynamics, weather and structural analysis. Code is published on GitHub under the Apache 2.0 licence.

Simulation · Analytics & Decision Intelligence

What it does

How this product is used

Capabilities

Not specified

Sectors

Cross-industry / not specified

Deployment

Not specified

Autonomy

ASSISTIVE

Data exposure

Not specified

Category context

Why buyers look at Simulation

Software that models queues, throughput and markets so operators can try a change without touching production. Jobs: queueing and throughput trials; synthetic counterparties in markets; what-if on operational flow.

Teams cannot rehearse a process without touching production.

This is category context, not an independent assessment of NVIDIA PhysicsNeMo. Vendor-claimed differences are listed above. Assessment results appear only after examination.

Browse Simulation →

Where to get more information

• Product website · developer.nvidia.com/physicsnemo
• NVIDIA website
• Company listing
• Ask to be invited to review
• Claim this listing to add a supplier statement

Other products from NVIDIA

• Canary-1b-v2 Claim

  Canary-1b-v2 is an open NVIDIA NeMo speech model with 1 billion parameters for speech transcription and translation across 25 European languages, including English, French, German, Spanish, Polish, Ukrainian and Russian. It performs both automatic speech recognition and speech translation, is available on Hugging Face with a demo, and is released under the CC-BY-4.0 licence for commercial and non-commercial use.

• Canary-Qwen-2.5B Claim

  Canary-Qwen-2.5B is an open NVIDIA NeMo English speech recognition model with 2.5 billion parameters, running at 418 RTFx and producing transcripts with punctuation and capitalisation. It works in two modes: an ASR mode that only transcribes, and an LLM mode that keeps the language model's abilities to post-process a transcript, for example to summarise it or answer questions about it. It is ready for commercial use under the CC-BY-4.0 licence.

• ChatRTX Claim

  NVIDIA sells GPUs, CUDA software and AI enterprise stacks used for training and inference. Public pages cover data-centre, cloud and on-prem AI compute. ChatRTX is NVIDIA's software product, from NVIDIA Corporation. Published at nvidia.com. Directory filing: General-purpose assistants. Vendor legal name on this product record: NVIDIA Corporation. ChatRTX is NVIDIA's software product, from NVIDIA Corporation. Published at nvidia.com. Directory filing: General-purpose assistants.

• FoundationPose Claim

  NVIDIA sells GPUs, CUDA software and AI enterprise stacks used for training and inference. Public pages cover data-centre, cloud and on-prem AI compute. FoundationPose is NVIDIA's model API, from NVIDIA Corporation. Published at build.nvidia.com/nvidia/foundationpose. Directory filing: Spatial intelligence. Vendor legal name on this product record: NVIDIA Corporation. FoundationPose is NVIDIA's model API, from NVIDIA Corporation. Published at build.nvidia.com/nvidia/foundationpose. Directory filing: Spatial intelligence.

• garak Claim

  garak (Generative AI Red-teaming and Assessment Kit) is a free, open-source command-line LLM vulnerability scanner maintained in NVIDIA's GitHub organisation. It checks whether a large language model or dialogue system can be made to fail in unwanted ways, probing for hallucination, data leakage, prompt injection, misinformation, toxicity generation, jailbreaks and other weaknesses using a mix of static, dynamic and adaptive probes. It works with Hugging Face Hub models, Replicate, the OpenAI API, AWS Bedrock, LiteLLM, GGUF models via llama.cpp and most models reachable over REST. garak is released under the Apache 2.0 licence and installs from PyPI with pip.

• Isaac ROS nvblox Claim

  Isaac ROS Nvblox is a set of NVIDIA ROS 2 packages for real-time 3D scene reconstruction and navigation cost maps. It takes depth and pose input from depth cameras and/or 3D LiDAR, builds a 3D reconstruction of the scene on the GPU and outputs a 2D costmap for Nav2, giving mobile robots a vision-based way to avoid obstacles while planning routes. Pose can come from Isaac ROS Visual SLAM or another pose estimation node. The packages are built on nvblox, NVIDIA's framework-independent, GPU-accelerated C++ library, and include examples for RealSense, ZED and Ouster LiDAR sensors as part of the wider Isaac ROS collection.

• Jetson AGX Orin Developer Kit Claim

  The NVIDIA Jetson AGX Orin Developer Kit is a compact edge AI computer for prototyping generative AI, robotics and computer vision products. It delivers up to 275 TOPS from a 2048-core NVIDIA Ampere architecture GPU with 64 Tensor Cores and a 12-core Arm Cortex-A78AE CPU, with 64GB of LPDDR5 memory, 64GB eMMC storage and power configurable from 15W to 60W. Because it shares one system-on-chip architecture with every Jetson Orin module, it can emulate any of them, so teams can start development before choosing a production module. I/O includes 10GbE, DisplayPort 1.4a, USB 3.2, PCIe Gen4 slots, a 16-lane MIPI CSI-2 camera connector and a 40-pin header, and it runs NVIDIA JetPack, Isaac ROS and the wider Jetson software stack.

• Jetson AGX Orin Industrial module Claim

  Jetson AGX Orin Industrial is the ruggedised production module in NVIDIA's Jetson AGX Orin series, built for edge AI and robotics products that need industrial-grade hardware. It delivers up to 248 TOPS from a 2048-core NVIDIA Ampere architecture GPU with 64 Tensor Cores and a 12-core Arm Cortex-A78AE CPU, with 64GB of 256-bit LPDDR5 memory with ECC and power configurable from 15W to 75W. It uses the same 100mm x 87mm, 699-pin form factor as the other AGX Orin modules, includes deep learning and vision accelerators, and supports up to six cameras (16 via virtual channels), PCIe Gen4, 10GbE and USB 3.2. It runs the same NVIDIA JetPack and Jetson software stack as the rest of the Orin family.

• Jetson Orin Nano module Claim

  Jetson Orin Nano modules are NVIDIA's entry-level production modules in the Jetson Orin family, sold in 8GB and 4GB versions in the small 69.6mm x 45mm, 260-pin SO-DIMM form factor. NVIDIA's current specifications list up to 67 TOPS for the 8GB module and 34 TOPS for the 4GB module, from an NVIDIA Ampere architecture GPU (1024 or 512 cores) and a 6-core Arm Cortex-A78AE CPU, with power modes between 7W and 25W. They support up to four cameras (eight via virtual channels), PCIe Gen3, USB 3.2 and gigabit Ethernet, with external NVMe storage. The modules run the same NVIDIA JetPack and Jetson software as larger Orin modules, and are aimed at drones, small robots and vision AI devices.

• Jetson Orin Nano Super Developer Kit Claim

  The NVIDIA Jetson Orin Nano Super Developer Kit is a compact, low-cost computer for running generative AI models at the edge, aimed at developers, students and makers. It delivers up to 67 TOPS of AI performance, a 1.7x improvement over the earlier Jetson Orin Nano Developer Kit, with 102 GB/s memory bandwidth and a 1.7 GHz CPU, and NVIDIA says it runs vision transformers, large language models and vision-language models. Existing Jetson Orin Nano Developer Kit owners can get the same boost through a JetPack software update. The kit includes a reference carrier board compatible with all Orin NX and Orin Nano modules, and is sold through NVIDIA's authorised distributors worldwide.

• Jetson Orin NX module Claim

  Jetson Orin NX modules are mid-range production modules in NVIDIA's Jetson Orin family, available in 16GB and 8GB versions in the small 69.6mm x 45mm, 260-pin SO-DIMM form factor. NVIDIA's current specifications list up to 157 TOPS for the 16GB module and 117 TOPS for the 8GB module, from a 1024-core NVIDIA Ampere architecture GPU with 32 Tensor Cores and an 8-core (16GB) or 6-core (8GB) Arm Cortex-A78AE CPU, with power modes from 10W to 40W. They include deep learning accelerators and support up to four cameras (eight via virtual channels), PCIe Gen4, USB 3.2 and gigabit Ethernet. The modules run NVIDIA JetPack, Isaac ROS and Metropolis software for robotics and edge AI products.

• Nemotron-3-Embed-1B-BF16 Claim

  Nemotron-3-Embed-1B-BF16 is a multilingual text embedding model from NVIDIA, optimised for retrieval and semantic similarity tasks and designed as a building block for text-based retrieval-augmented generation (RAG) systems. It turns questions and passages into dense vector embeddings for retrieval, semantic search and agentic RAG workflows, and NVIDIA evaluated it across 34 languages with cross-lingual retrieval. It is ready for commercial use under the OpenMDW 1.1 licence, was released on Hugging Face in July 2026 and sits alongside an NVFP4 1B variant and a larger 8B model in the Nemotron 3 Embed family.

• Nemotron-Mini-4B-Instruct Claim

  Nemotron-Mini-4B-Instruct is a small language model from NVIDIA for generating responses for roleplay, retrieval-augmented generation and function calling. It is optimised through distillation, pruning and quantisation for speed and on-device deployment, and is a fine-tuned version of NVIDIA's Minitron-4B-Base. NVIDIA uses it in NVIDIA ACE for in-game character roleplay. The model is a Nemotron-4 transformer decoder with Grouped-Query Attention and Rotary Position Embeddings, is published on Hugging Face under the NVIDIA Community Model License, and can also be tried on build.nvidia.com.

• NVIDIA ACE Claim

  NVIDIA ACE is a suite of AI technologies that helps game and middleware developers build conversational, autonomous in-game characters and AI assistants. It provides ready-to-integrate cloud and on-device models for speech recognition, text-to-speech, language and facial animation, optimised to run alongside graphics on gaming hardware with low latency and a small memory footprint. Developers can use MIT-licensed ACE plugins for Unreal Engine 5, including Audio2Face-3D facial animation, speech recognition, an on-device LLM and text-to-speech, or the Game Agent SDK and In-Game Inferencing SDK for custom engines. NVIDIA cites games such as KRAFTON's PUBG and inZOI, Total War: PHARAOH and Wemade Next's MIR5 as using ACE-powered characters.

• NVIDIA AI Enterprise Claim

  NVIDIA AI Enterprise is a commercial, fully supported software suite that brings together microservices, frameworks and libraries for AI development with GPU orchestration and infrastructure management for running AI in production. It includes ready-to-deploy NVIDIA NIM microservices and NVIDIA Blueprints, validated deployment guides, extended-lifetime production branches and enterprise support, along with a secure software supply chain and vulnerability mitigation. It runs on NVIDIA-Certified data centre and edge systems, uses the same software components from development to production, and integrates with IT management and MLOps tools, with centralised policy, governance and visibility for scaling workloads.

• NVIDIA Base Command Manager Claim

  NVIDIA Base Command Manager is cluster management software for provisioning, managing and monitoring high-performance computing (HPC) and AI clusters, from a few nodes to hundreds of thousands, across edge, data centre and multi- or hybrid-cloud deployments. It handles mixed, hardware-agnostic environments with GPU-accelerated, Arm and x86 nodes, auto-scales capacity by powering nodes up and down, and integrates with workload managers such as Slurm and NVIDIA Run:ai for bare-metal and containerised workloads. Administrators work through the BaseView web interface, which includes pre-built monitoring dashboards and installation wizards, or through a full command line and API. Clusters on Demand can spin complete clusters up and down on major cloud providers. Base Command Manager is available with a free licence, with optional enterprise support or entitlement through NVIDIA AI Enterprise and Mission Control.

• NVIDIA Broadcast Claim

  NVIDIA Broadcast is a desktop app for PCs with NVIDIA RTX GPUs that uses AI to improve voice and video for livestreams, voice chats and video calls. Audio features include Studio Voice and noise and room echo removal. Video features include background removal, replacement and blur, Virtual Key Light face relighting, Eye Contact, Auto Frame, and beta RTX Video Super Resolution and frame generation for webcams. A file processor applies the same effects to recordings before upload, and an MCP server lets AI agents change settings and process files.

• NVIDIA Cosmos3 Nano Claim

  NVIDIA Cosmos3 Nano is a world foundation model in NVIDIA's Cosmos family that generates images, videos and action predictions from text, visual inputs and spatial controls. Robotics and physical AI teams use it to produce synthetic scenes and predict how a scene will change. NVIDIA lists input screening and visual SynthID watermarking as built-in safeguards, and the model can be tried on the NVIDIA API catalogue.

• NVIDIA Cosmos Curator Claim

  NVIDIA Cosmos Curator is a toolkit for curating physical AI training data, built to process large video datasets into clean, captioned clips for training world foundation models. Robotics and autonomous vehicle teams use it to filter, split and annotate footage at scale on NVIDIA GPUs. Code is published on GitHub under the Apache 2.0 licence, and a container is available on NVIDIA NGC.

• NVIDIA Cosmos Predict 2.5 Claim

  Cosmos Predict 2.5 is a world foundation model in NVIDIA's Cosmos family that simulates and predicts the future state of a scene as video. Robotics and autonomous vehicle teams use it to generate future frames from text, image or video inputs for synthetic training data and policy evaluation. Code and post-training scripts are published on GitHub under the Apache 2.0 licence.

• NVIDIA Cosmos Reason 2 Claim

  Cosmos Reason 2 is an NVIDIA vision language model for physical AI that understands physical common sense and produces embodied decisions in natural language through long chain-of-thought reasoning. Robotics and video analytics teams use it to interpret scenes, critique synthetic data and plan actions. Code is published on GitHub under the Apache 2.0 licence.

• NVIDIA Cosmos-Tokenize1-CV8x8x8-720p Claim

  Cosmos-Tokenize1-CV8x8x8-720p is a continuous video tokenizer from NVIDIA's Cosmos Tokenizer suite, which compresses images and videos into compact latent representations while keeping reconstruction quality high. This variant compresses video by eight times in time and in each spatial dimension and targets 720p footage. Developers use it as the encoding stage for video generation and world models. Weights are published on Hugging Face.

• NVIDIA Cosmos Transfer 2.5 Claim

  Cosmos Transfer 2.5 is an NVIDIA world foundation model built on Cosmos Predict 2.5 that generates photorealistic world simulations conditioned on spatial control inputs such as depth, segmentation and edge maps. Teams feed it simulator output, for example from Omniverse, to produce varied, realistic synthetic video for training physical AI systems. Code is on GitHub under the Apache 2.0 licence.

• NVIDIA CUDA-X AI Claim

  NVIDIA CUDA-X is a collection of hundreds of GPU-accelerated libraries built on the CUDA platform that give developers ready-made algorithms and computational routines for AI, data processing, high-performance computing, scientific computing and quantum computing, without low-level GPU programming. Deep learning libraries include cuDNN, TensorRT and TensorRT-LLM, plus CUTLASS and FlashInfer for custom kernels. Data processing libraries speed up pandas, Polars, scikit-learn and Apache Spark without code changes and add cuGraph, cuVS and cuOpt for graph analytics, vector search and optimisation. Other libraries cover image and video processing (DALI and CV-CUDA), multi-GPU communication (NCCL, NVSHMEM and NIXL), parallel algorithms (Thrust and CUB), physics-aware AI (PhysicsNeMo and Warp) and quantum simulation and post-quantum cryptography (cuPQC).

• NVIDIA cuOpt Claim

  NVIDIA cuOpt is an open-source, GPU-accelerated decision optimisation engine for large problems with millions of variables and constraints. It supports linear programming, vehicle routing and quadratic programming, with beta support for mixed integer programming, quadratically constrained quadratic programming and second-order cone programming. Existing AMPL, CVXPY, PuLP and JuMP models can be accelerated with minimal code changes, models can be rerun in near real time or in batch, and it scales across hybrid and multi-cloud environments. NVIDIA-verified agent skills help AI agents formulate, solve, debug and explain optimisation problems, and enterprise support is available through NVIDIA AI Enterprise.

• NVIDIA DeepStream SDK Claim

  NVIDIA DeepStream SDK is an open-source, GStreamer-based toolkit for real-time streaming analytics, covering AI-based multi-sensor processing and video, audio and image understanding. It is aimed at developers, software partners, startups and OEMs building vision AI agents and applications for areas such as smart cities, retail, manufacturing and logistics, and can be deployed on premises, at the edge or in the cloud. DeepStream ships with more than 40 hardware-accelerated plug-ins and over 30 sample applications, supports C/C++ and Python, and offers agent skills so coding agents can generate video analytics pipelines from natural-language prompts. Service Maker simplifies building C++ pipelines, and the Multiview 3D tracking extension follows objects across networks of cameras. DeepStream is part of NVIDIA Metropolis.

• NVIDIA DGX Cloud Claim

  NVIDIA now describes DGX Cloud as its internal cloud environment for building and operating AI at scale. It is used to develop open frontier and foundation models such as the NVIDIA Nemotron family, validate new system architectures and run production AI workloads, and it runs on NVIDIA-accelerated infrastructure across cloud service providers and NVIDIA Cloud Partners. The software and operating patterns proven in DGX Cloud are released to partners as open, modular components through NVIDIA DSX OS, the operating layer for AI factories, while NVIDIA Exemplar Cloud provides validated hardware and software stacks with performance benchmarking.

• NVIDIA DRIVE Claim

  NVIDIA DRIVE is NVIDIA's platform for building autonomous vehicles and robotaxis. It brings together DRIVE AGX in-vehicle AI computers, the DriveOS safety-certified operating system, the Hyperion sensor and compute reference platform for Level 4 vehicles, high-fidelity sensor simulation, open autonomous-driving models and datasets, and data-centre infrastructure for training and validating driving systems at scale, with a safety architecture covering the development lifecycle.

• NVIDIA DRIVE AGX Orin Claim

  The NVIDIA DRIVE AGX Orin Developer Kit is a bench platform for developing production-level autonomous vehicle software, with a vehicle accessory kit available separately for in-vehicle work. It is built on a single automotive-grade AGX Orin system-on-chip with an Ampere architecture-class GPU delivering up to 254 INT8 TOPS, including a deep learning accelerator, plus 12 Arm Cortex-A78A CPU cores, up to 200 GB/s memory bandwidth and 256GB UFS storage. Automotive I/O includes 16 GMSL camera ports, 10GbE and 1GbE Ethernet for lidar and radar data, and six CAN interfaces, and several kits can be linked for more performance. It runs NVIDIA DriveOS with DriveWorks plus NvMedia, CUDA, cuDNN and TensorRT.

• NVIDIA DRIVE AGX Thor Claim

  The NVIDIA DRIVE AGX Thor Developer Kit gives automotive teams the hardware, software and sample applications to build production-level autonomous vehicle and in-car AI applications. It is built on a single AGX Thor system-on-chip with a Blackwell-class GPU delivering up to 2,000 FP4 (1,000 INT8) TFLOPS, an Arm Neoverse V3AE CPU, 64GB of LPDDR5X memory and 256GB UFS storage, on automotive-grade silicon designed for safety and security. Automotive I/O includes 16 GMSL2 plus 2 GMSL3 camera links, multi-gigabit automotive Ethernet and four CAN interfaces, and multiple units can be linked for more performance. It ships in a bench version (SKU 10) and an in-vehicle version (SKU 12) and runs NVIDIA DriveOS with DriveWorks, CUDA, cuDNN and TensorRT.

• NVIDIA DRIVE Hyperion Claim

  NVIDIA sells GPUs, CUDA software and AI enterprise stacks used for training and inference. Public pages cover data-centre, cloud and on-prem AI compute. NVIDIA DRIVE Hyperion is NVIDIA's robotics product, from NVIDIA Corporation. Published at developer.nvidia.com/drive/hyperion. Directory filing: Autonomous vehicles. Vendor legal name on this product record: NVIDIA Corporation. NVIDIA DRIVE Hyperion is NVIDIA's robotics product, from NVIDIA Corporation. Published at developer.nvidia.com/drive/hyperion. Directory filing: Autonomous vehicles.

• NVIDIA DriveOS Claim

  NVIDIA DriveOS is an automotive operating system and software development kit for NVIDIA DRIVE AGX hardware, developed with safety and security methods certified by TÜV SÜD. It runs Linux or QNX as the application operating system and provides tools for image processing, sensor integration, AI acceleration with CUDA and TensorRT, and low-overhead communication between processes, including NvMedia for loading camera frames straight into GPU memory and NvStreams for zero-copy data transfer between hardware accelerators. DriveOS is designed for safety-critical applications and complies with ASPICE, ISO 26262 and ISO/SAE 21434. A hypervisor isolates guest operating systems for redundancy, Docker containers are supported on host and target, and NVIDIA Nsight handles debugging and profiling. The SDK is downloaded through the DRIVE AGX SDK Developer Program and includes DriveWorks, CUDA, cuDNN and TensorRT.

• NVIDIA Dynamo Claim

  NVIDIA Dynamo is an open-source, low-latency, modular inference framework for serving generative AI models in distributed environments. It scales inference across large GPU fleets by disaggregating the prefill and decode phases across GPUs, routing requests to avoid redundant computation and extending GPU memory by caching data to lower-cost storage tiers. It works with open-source inference engines including SGLang, TensorRT LLM and vLLM. Key components include an SLO Planner that adjusts GPU resources to meet service level objectives, a KV-cache-aware router, the NIXL low-latency data transfer library, a KV Block Manager and Grove for multi-node deployment on Kubernetes. Dynamo builds on NVIDIA Triton Inference Server.

• NVIDIA Earth-2 Claim

  NVIDIA Earth-2 is a family of open models, libraries and frameworks for weather and climate AI, covering every stage of forecasting from processing observation data to high-resolution visualisation. It is intended to let researchers, startups and government agencies run, fine-tune and deploy forecasting systems on their own infrastructure. Models include Earth-2 Medium Range, built on the Atlas architecture, which NVIDIA says predicts over 70 weather variables up to 15 days ahead, and Earth-2 Nowcasting, which uses generative AI to predict satellite and radar imagery for local storm forecasts.

• NVIDIA Holoscan Claim

  NVIDIA Holoscan is a domain-agnostic, multimodal AI sensor processing platform that provides accelerated, full-stack infrastructure for real-time processing of streaming data at the edge or in the cloud. The open-source Holoscan SDK can be installed via pip, Conda, a Debian package or an NGC container, and runs on industrial-grade hardware such as NVIDIA IGX as well as other aarch64 and x86 platforms. Holoscan Sensor Bridge brings high-bandwidth sensor data in over Ethernet through an FPGA interface with a standard API. The HoloHub repository shares reference applications and operators, including an end-to-end surgical video workflow, 3D Slicer integration, XR volume rendering, body pose estimation and SAM2 segmentation on live video.

• NVIDIA IGX Claim

  NVIDIA IGX is an industrial-grade edge AI platform that combines enterprise hardware, software and support for industrial, robotics and medical environments, with high-bandwidth sensor processing, enterprise security and functional safety. The current generation, IGX Thor, is built on the NVIDIA Blackwell architecture with an integrated GPU and an optional discrete GPU, delivering up to 5,581 FP4 TFLOPS, and NVIDIA says it provides up to eight times the integrated-GPU AI compute of IGX Orin. A dedicated Functional Safety Island isolates safety-critical workloads for robots and machines working alongside people. IGX runs NVIDIA AI Enterprise software, including NIM microservices and models such as Isaac GR00T N and Cosmos Reason, and is available as developer kits, a system-on-module for custom carrier boards and certified production systems from OEM partners.

• NVIDIA IGX Orin Developer Kit Claim

  The NVIDIA IGX Orin Developer Kit is a development system for industrial and medical edge AI, part of NVIDIA's IGX industrial-grade platform for industrial, robotics and medical environments. It comes complete with chassis and power supply and is built on an IGX Orin module with a 2,048-core Ampere architecture GPU with 64 Tensor Cores and a 12-core Arm Cortex-A78AE CPU, reaching up to 1705 TOPS with an optional RTX 6000 Ada Generation GPU. It adds NVIDIA ConnectX-7 with 2x 100GbE networking, a safety microcontroller and a baseboard management controller. The IGX platform comes with NVIDIA AI Enterprise and up to 10 years of support; NVIDIA now also sells the newer, Blackwell-based IGX Thor alongside it.

• NVIDIA Isaac Claim

  NVIDIA Isaac is an open robotics development platform made up of simulation and robot learning frameworks, CUDA-accelerated libraries, AI models and reference workflows for building autonomous mobile robots, robot arms, manipulators and humanoids. It includes Isaac ROS, an open-source foundation for building robots with ROS 2; Isaac Manipulator for AI-powered robotic arms; and Isaac GR00T, a development platform and set of robot foundation models for humanoid robotics. Isaac Sim, built on NVIDIA Omniverse, lets teams develop and test robots in a physically based virtual environment and, with NVIDIA Cosmos, generate synthetic training data, while Isaac Lab is optimised for robot learning and foundation model training. NVIDIA pairs the software with DGX systems for training, OVX for simulation and AGX systems such as Jetson for deployment.

• NVIDIA Isaac GR00T N1 Claim

  Isaac GR00T N is NVIDIA's family of open robot foundation models for general-purpose humanoid robots, first released as GR00T N1 and now available as Isaac GR00T 1.7 in early access on Hugging Face. The models take a video sequence from onboard cameras, a natural-language command and the robot's current joint positions, and output short sequences of joint motions for smooth whole-body control, locomotion and two-handed manipulation such as grasping, moving objects and passing items between arms. They are trained on a mix of internet-scale human video, real-world teleoperation and synthetic data, and can be adapted to new robot hardware by fine-tuning on a small set of recorded demonstrations. GR00T N sits within the wider Isaac GR00T platform alongside Isaac Lab, Isaac Sim, teleoperation and whole-body control tools, with NVIDIA Jetson Thor for on-robot inference.

• NVIDIA Isaac Lab Claim

  NVIDIA Isaac Lab is an open-source, GPU-accelerated simulation framework for robot learning, designed to train robot policies at scale. Its modular architecture supports humanoids, manipulators and autonomous mobile robots, and covers both reinforcement learning and imitation learning, with custom libraries such as skrl, RLLib and rl_games. Teams can choose physics engines including Newton, NVIDIA PhysX, NVIDIA Warp and MuJoCo, use tiled rendering to feed multi-camera vision data into training, and scale across multiple GPUs and nodes locally or on AWS, GCP, Azure and Alibaba Cloud through NVIDIA OSMO. Isaac Lab is the robot learning framework behind the NVIDIA Isaac GR00T platform, and Isaac Lab-Arena adds policy evaluation in simulation.

• NVIDIA Isaac Manipulator Claim

  NVIDIA sells GPUs, CUDA software and AI enterprise stacks used for training and inference. Public pages cover data-centre, cloud and on-prem AI compute. NVIDIA Isaac Manipulator is NVIDIA's platform, from NVIDIA Corporation. Published at developer.nvidia.com/isaac/manipulator. Directory filing: Physical AI. Vendor legal name on this product record: NVIDIA Corporation. NVIDIA Isaac Manipulator is NVIDIA's platform, from NVIDIA Corporation. Published at developer.nvidia.com/isaac/manipulator. Directory filing: Physical AI.

• NVIDIA Isaac Perceptor Claim

  NVIDIA sells GPUs, CUDA software and AI enterprise stacks used for training and inference. Public pages cover data-centre, cloud and on-prem AI compute. NVIDIA Isaac Perceptor is NVIDIA's platform, from NVIDIA Corporation. Published at developer.nvidia.com/isaac/perceptor. Directory filing: Physical AI. Vendor legal name on this product record: NVIDIA Corporation. NVIDIA Isaac Perceptor is NVIDIA's platform, from NVIDIA Corporation. Published at developer.nvidia.com/isaac/perceptor. Directory filing: Physical AI.

• NVIDIA Isaac ROS Claim

  NVIDIA Isaac ROS is an open-source software foundation for building AI-powered robots with ROS 2. Built on NVIDIA's upstream contributions to ROS and compatible with open ROS standards, it adds GPU-accelerated packages, AI models, libraries, workflows and deployment tools so developers can keep familiar ROS 2 tools while moving from prototyping in Isaac Sim to deployment on NVIDIA Jetson. NVIDIA Isaac Transport for ROS (NITROS) speeds up ROS 2 processing pipelines while keeping portability and interoperability, and modular perception packages slot into existing ROS 2 applications. NVIDIA also provides agent skills, for example to help CUDA-accelerate an existing ROS 2 node.

• NVIDIA Isaac Sim Claim

  NVIDIA Isaac Sim is an open-source reference framework, built on NVIDIA Omniverse libraries, for robotics simulation, testing and synthetic data generation in physically based virtual environments. It ingests CAD, URDF robot descriptions or real-world captures and converts them to OpenUSD, so teams can assemble scenes with materials, physics, robot and sensor models. Robots can be trained with Isaac Lab, synthetic data pipelines can be augmented with NVIDIA Cosmos world foundation models, and full perception and mobility stacks can be evaluated with software-in-the-loop or hardware-in-the-loop testing. It is extensible for custom OpenUSD simulators and is available from GitHub or on Brev.

• NVIDIA Jetson Claim

  NVIDIA Jetson is a platform of compact edge AI computers for robotics and edge AI applications, supported by the NVIDIA JetPack SDK. JetPack provides pre-built, cloud-native software services for developing and deploying AI inference at the edge, including generative AI, computer vision, advanced robotics and space computing. The hardware range spans the Blackwell-based Jetson Thor series for high-end physical AI, the Jetson Orin series and Orin Nano Super developer kit, and older Xavier, TX2 and Nano modules. NVIDIA has also announced Jetson Orin Nano 2, an entry-level robotics computer for drones, robots and vision AI systems.

• NVIDIA Jetson T4000 Claim

  The NVIDIA Jetson T4000 is a Jetson Thor series production module that brings Blackwell-based AI to a wider range of robotics and edge AI products. It delivers up to 1200 FP4 (sparse) TFLOPS from a 1536-core Blackwell architecture GPU with fifth-generation Tensor Cores and Multi-Instance GPU support, with a 12-core Arm Neoverse-V3AE CPU and 64GB of LPDDR5X memory, at 40W to 70W. It shares its form factor and pin layout with the Jetson T5000, so one carrier board design can serve both, and NVIDIA reports up to 2x the performance of Jetson AGX Orin on language, speech and vision-language-action workloads. It launched in January 2026 alongside JetPack 7.1, which added the TensorRT Edge-LLM SDK.

• NVIDIA Jetson T5000 Claim

  The NVIDIA Jetson T5000 is the top production module in the Jetson Thor series, built for humanoid robots, physical AI and agentic AI at the edge. It delivers up to 2070 FP4 (sparse) TFLOPS from a Blackwell architecture GPU with fifth-generation Tensor Cores and Multi-Instance GPU support, with a 14-core Arm Neoverse-V3AE CPU and 128GB of LPDDR5X memory at 273 GB/s, at 40W to 130W. NVIDIA quotes 7.5x the AI performance and 3.5x the energy efficiency of Jetson AGX Orin. Networking includes 4x 25GbE for high-speed sensor input, and it supports up to 20 cameras through Holoscan Sensor Bridge. It runs NVIDIA Isaac, GR00T, Holoscan and video search and summarisation workflows.

• NVIDIA Jetson Thor Claim

  NVIDIA Jetson Thor is a series of Blackwell-powered compute modules and developer kits for physical AI, humanoid robots and edge AI. NVIDIA quotes up to 2,070 FP4 TFLOPS of AI compute, 128 GB of memory, 273 GB/s memory bandwidth and a 14-core CPU at 40 to 130 W, which it says is 7.5 times the AI performance and 3.5 times the energy efficiency of Jetson AGX Orin. It targets generative AI and large transformer models at the edge, high-speed sensor processing and real-time reasoning, supports popular open models through Jetson AI Lab, and adds Jetson agent skills for deploying autonomous agents. Newer T3000 and T2000 modules extend the range, and units are bought through NVIDIA partners.

• NVIDIA MCG Toolkit Claim

  The NVIDIA model card generator (MCG) toolkit automates AI model documentation by reading a model's source code and producing a model card in NVIDIA's Model Card++ format, usually in under a minute. It accepts a GitHub, GitLab or Hugging Face URL, any public web page, or an uploaded ZIP, PDF, DOCX or Markdown file, through a web interface or REST API. A containerised ingestion, extraction and rendering pipeline uses NVIDIA Nemotron RAG embedding and reranking models with GPT-OSS-120B to fill in an overview plus Bias, Explainability, Privacy and Safety & Security subcards, and outputs are CycloneDX-compliant. Models, templates and field guidance are configurable, and it runs on-premises or in your own cloud with Kubernetes support, aimed at teams facing rules such as the EU AI Act.

• NVIDIA Merlin Claim

  NVIDIA Merlin is an open-source framework for building high-performing recommender systems at scale. It gives data scientists, machine learning engineers and researchers libraries, methods and tools for preprocessing, feature engineering, training, inference and deployment to production, optimised for retrieval, filtering, scoring and ordering across hundreds of terabytes of data through straightforward APIs. Components such as Merlin Models, which provides standard recommender models, are designed to interoperate with existing data science, machine learning and deep learning workflows on CPUs or GPUs, so teams can adopt one library or the whole pipeline.

• NVIDIA Metropolis Claim

  NVIDIA Metropolis is a vision AI application platform and partner ecosystem for developing, deploying and scaling visual AI agents from the edge to the cloud. It provides vision language and computer vision models plus microservices for video ingestion, storage and model tuning, so organisations can search and understand live or archived video using natural language. Use cases include video analytics agents for warehouses, retail, airports and traffic intersections, automated visual inspection for manufacturing defects, intelligent transportation with multi-camera tracking, robot fleet management through the Mega Omniverse Blueprint, and loss prevention and stock visibility in retail stores.

• NVIDIA MONAI Claim

  MONAI is an open-source, PyTorch-based framework for medical imaging AI, promoted by NVIDIA as part of its Clara healthcare platform. It brings together domain-optimised data loaders, transforms and more than 50 pretrained models to support 2D and 3D segmentation, registration, annotation, model training and reporting. MONAI also supports multimodal workflows that combine imaging, text and agentic AI for radiology, surgery and pathology. NVIDIA reports more than 8 million downloads and over 4,000 projects built on it, with code available on GitHub.

• NVIDIA NeMo Claim

  NVIDIA NeMo is an agent-first, open suite of libraries and skills for building, specialising, optimising and governing AI agents across cloud, on-premises and hybrid environments, and it integrates with existing AI tools and agent frameworks. It covers the agent lifecycle: NeMo Curator, Data Designer, Anonymizer and Safe Synthesizer prepare real and synthetic data; NeMo Evaluator benchmarks models and agents; NeMo Customizer, NeMo RL and NeMo Gym fine-tune and post-train models; NeMo Guardrails and NeMo Auditor handle safety, compliance and vulnerability testing; and NeMo Relay connects and observes external agent harnesses. NeMo works alongside NVIDIA Nemotron models and NIM microservices for deployment, and the open-source NeMo Framework supports training and aligning LLMs and multimodal models at scale.

• NVIDIA NeMo Curator Claim

  NVIDIA NeMo Curator is an open-source platform for scalable, privacy-aware curation of text, image, video and audio data for training large language models and generative AI. Text pipelines run on Ray and cover download, extraction, language detection, rule-based and model-based quality filtering and exact, fuzzy and semantic deduplication. Image pipelines generate embeddings, classify for aesthetics and NSFW content, filter and export; video pipelines split long videos into clips with optional encoding, filtering, embedding and deduplication; and audio pipelines run speech recognition and score quality by word error rate. It uses GPUs, Ray and NVIDIA RAPIDS to scale from a single laptop to a multi-node GPU cluster, runs in the cloud, on-premises or hybrid, and is part of the NVIDIA NeMo suite.

• NVIDIA NeMo Customizer Claim

  NVIDIA sells GPUs, CUDA software and AI enterprise stacks used for training and inference. Public pages cover data-centre, cloud and on-prem AI compute. NVIDIA NeMo Customizer is NVIDIA's platform, from NVIDIA Corporation. Published at developer.nvidia.com/nemo-customizer. Directory filing: Fine-tuning platforms. Vendor legal name on this product record: NVIDIA Corporation. NVIDIA NeMo Customizer is NVIDIA's platform, from NVIDIA Corporation. Published at developer.nvidia.com/nemo-customizer. Directory filing: Fine-tuning platforms.

• NVIDIA NeMo Data Designer Claim

  NVIDIA NeMo Data Designer is a compound AI system for generating synthetic datasets, part of the NVIDIA NeMo suite. Users define the columns of a dataset in Python, mixing sampled values such as categories and personas with columns generated by a language model from templated prompts, and Data Designer handles output formatting for each column type and returns the results as a pandas DataFrame. A hosted free trial on build.nvidia.com needs only an NVIDIA API key, uses the Python SDK and generates up to 100 records per job, with Nemotron, Mistral Small and Llama 4 Scout models available by default. The trial is governed by NVIDIA's API trial terms and the licence of each model used.

• NVIDIA NeMo Guardrails Claim

  NVIDIA NeMo Guardrails is an open-source Python library for adding programmable guardrails to LLM applications, blocking, altering or validating unsafe, off-topic, malicious or policy-violating inputs and model responses. Input, retrieval, dialog, execution and output rails run at different stages of an interaction, configured in YAML and Colang flows and extended with custom Python actions or external APIs. It supports LLM self-checking, NVIDIA NemoGuard safety models, community models such as LlamaGuard and third-party services such as ActiveFence and Cisco AI Defense, including jailbreak protection. The same configuration runs locally with the library or in production with the NeMo Guardrails microservice on Kubernetes.

• NVIDIA Nemotron Nano 12B v2 VL Claim

  Nemotron Nano 12B v2 VL is a 12-billion-parameter NVIDIA vision language model for multi-image and video understanding. It handles visual question answering and video summarisation, and is offered as an NVIDIA NIM on the API catalogue for hosted trials or self-hosted deployment on NVIDIA GPUs.

• NVIDIA NIM for Grounding DINO Claim

  NVIDIA sells GPUs, CUDA software and AI enterprise stacks used for training and inference. Public pages cover data-centre, cloud and on-prem AI compute. NVIDIA NIM for Grounding DINO is NVIDIA's model API, from NVIDIA Corporation. Published at build.nvidia.com/nvidia/grounding-dino. Directory filing: Object detection. NVIDIA NIM for Grounding DINO is NVIDIA's model API, from NVIDIA Corporation. Published at build.nvidia.com/nvidia/grounding-dino. Directory filing: Object detection.

• NVIDIA NIM for NV-CLIP Claim

  NVIDIA NIM for NV-CLIP packages NV-CLIP, a multimodal embeddings model for images and text, as a deployable inference microservice. Developers use it to embed images and text into a shared vector space for semantic image search, zero-shot classification and retrieval in multimodal applications. It can be tried on the NVIDIA API catalogue and self-hosted on NVIDIA GPUs.

• NVIDIA NIM for RetinaNet Claim

  NVIDIA sells GPUs, CUDA software and AI enterprise stacks used for training and inference. Public pages cover data-centre, cloud and on-prem AI compute. NVIDIA NIM for RetinaNet is NVIDIA's model API, from NVIDIA Corporation. Published at build.nvidia.com/nvidia/retinanet. Directory filing: Object detection. Vendor legal name on this product record: NVIDIA Corporation. NVIDIA NIM for RetinaNet is NVIDIA's model API, from NVIDIA Corporation. Published at build.nvidia.com/nvidia/retinanet. Directory filing: Object detection.

• NVIDIA NIM / NeMo Claim

  NVIDIA NIM and NeMo provide inference microservices and tooling to build, customise, and deploy AI models on NVIDIA infrastructure. Public marketing also emphasises gPU ecosystem standard. It is presented for Enterprises deploying models on NVIDIA stacks. NVIDIA NIM / NeMo is NVIDIA's platform, from NVIDIA Corporation. Published at nvidia.com. Directory filing: Model hosting / inference.

• NVIDIA Omniverse Claim

  NVIDIA Omniverse is a collection of libraries and services, part of NVIDIA Agent Toolkit, for building simulation-ready worlds that physical AI teams use to train, test and validate systems before real-world deployment. Developers integrate capabilities for OpenUSD data interoperability, RTX rendering, GPU-accelerated physics, sensor simulation, SimReady asset validation and runtime workflows into their own applications, and AI agents can call these as tools through Omniverse Agent Skills. Teams can start from libraries, agent skills, reference blueprints or experimental samples, with uses including industrial facility digital twins and robotics simulation.

• NVIDIA Omniverse for Digital Twins Claim

  NVIDIA sells GPUs, CUDA software and AI enterprise stacks used for training and inference. Public pages cover data-centre, cloud and on-prem AI compute. NVIDIA Omniverse for Digital Twins is NVIDIA's platform, from NVIDIA Corporation. Published at nvidia.com. Directory filing: Digital twins. Vendor legal name on this product record: NVIDIA Corporation. NVIDIA Omniverse for Digital Twins is NVIDIA's platform, from NVIDIA Corporation. Published at nvidia.com. Directory filing: Digital twins.

• NVIDIA Omniverse Replicator Claim

  Omniverse Replicator is an NVIDIA framework for building custom synthetic data generation pipelines and services. Developers use it to produce physically accurate 3D synthetic data to train and test perception networks for autonomous vehicles, robotics and intelligent video analytics. It runs as an extension within NVIDIA Omniverse.

• NVIDIA Omniverse Sensor RTX Claim

  NVIDIA sells GPUs, CUDA software and AI enterprise stacks used for training and inference. Public pages cover data-centre, cloud and on-prem AI compute. NVIDIA Omniverse Sensor RTX is NVIDIA's platform, from NVIDIA Corporation. Published at docs.omniverse.nvidia.com/sensors-rtx/latest/index.html. Directory filing: Synthetic environments. NVIDIA Omniverse Sensor RTX is NVIDIA's platform, from NVIDIA Corporation. Published at docs.omniverse.nvidia.com/sensors-rtx/latest/index.html. Directory filing: Synthetic environments.

• NVIDIA PhysicsNeMo DoMINO Automotive Aero NIM Claim

  NVIDIA PhysicsNeMo DoMINO Automotive Aero NIM packages the DoMINO model for automotive aerodynamics as a deployable inference microservice. NVIDIA presents it as a demonstration of foundation models for early-stage design evaluation, letting engineers estimate airflow around a vehicle shape far faster than a full simulation. The container is available on NVIDIA NGC with documentation on the NVIDIA docs site.

• NVIDIA PhysicsNeMo Framework Claim

  NVIDIA PhysicsNeMo is an open-source Python framework for building, training and fine-tuning physics-ML models at scale. Engineers use it to create neural surrogates for simulations such as computational fluid dynamics, structural mechanics and weather, combining physics-informed methods with neural operators and graph networks. It is released on GitHub under the Apache 2.0 licence.

• NVIDIA RAPIDS Claim

  RAPIDS is NVIDIA's suite of open-source, GPU-accelerated data science libraries, now presented as part of the CUDA-X data science libraries. cuDF accelerates DataFrame work with drop-in, no-code-change acceleration for pandas, Polars and Apache Spark; cuML runs machine learning algorithms from scikit-learn, UMAP and HDBSCAN on GPUs without code changes; and cuGraph accelerates graph analytics with a zero-code-change accelerator for NetworkX. cuxfilter adds interactive dashboards over very large tables, Dask scales pipelines across multiple nodes, and the RAPIDS Accelerator for Apache Spark speeds up Spark processing. The libraries install with conda or pip and deploy locally, on Kubernetes, Databricks or Google Colab, or on AWS, Azure and GCP.

• NVIDIA Riva Claim

  NVIDIA Riva is the library for optimised inference of NVIDIA Nemotron Speech, a family of GPU-accelerated open models for automatic speech recognition, text-to-speech, neural machine translation and speech-to-speech across around 40 languages. The models are pretrained and can be fine-tuned on custom data for domain-specific use, and are delivered as containerised microservices built for real-time performance and high-throughput offline work, scaling to thousands of parallel streams. They run across clouds, in data centres, at the edge and on embedded devices, and NVIDIA positions them for building voice-first AI agents, with NIM microservices for production voice experiences.

• NVIDIA TAO Toolkit Claim

  NVIDIA TAO is a suite of skills and tools for post-training vision language and vision foundation models, which can now be run as agentic workflows from natural-language prompts in a coding agent. Teams start from a pretrained model in the TAO model zoo or a supported third-party architecture, fine-tune it on their own domain, and use knowledge distillation and quantisation to make it smaller and faster before exporting it in open formats for deployment from edge to cloud with NVIDIA DeepStream. TAO 7 adds data sourcing skills that diagnose accuracy weaknesses and fill gaps with real or synthetic data, an AutoML agent for hyperparameter optimisation and annotation skills for dataset preparation. NVIDIA positions it for custom industrial AI such as visual inspection, quality control and robotic guidance.

• NVIDIA TensorRT Claim

  NVIDIA TensorRT is an ecosystem of tools for high-performance deep learning inference, including inference compilers, runtimes and model optimisations for low latency and high throughput in production. Built on CUDA, it optimises models trained in all major frameworks, calibrates them for lower precision and deploys them to data centres, workstations, laptops and edge devices, using quantisation, layer and tensor fusion and kernel tuning; NVIDIA quotes up to 36 times faster inference than CPU-only platforms. The family includes the TensorRT compiler, TensorRT-LLM, TensorRT Model Optimizer (FP8, FP4, INT8, INT4 and AWQ quantisation), TensorRT for RTX and TensorRT Cloud.

• NVIDIA TensorRT-LLM Claim

  NVIDIA TensorRT LLM is an open-source library for high-performance, real-time inference of large language models on NVIDIA GPUs, on a desktop or in a data centre. It offers a modular Python runtime, PyTorch-native model authoring, a simple CLI and a stable production LLM API, with kernels designed for NVIDIA hardware. Optimisations include FP8 and NVFP4 quantisation, disaggregated serving, parallelisation including wide expert parallelism, and speculative decoding techniques such as EAGLE-3 and multi-token prediction, aimed at serving more users in parallel at lower operating cost. It is available from NVIDIA NGC and GitHub.

• NVIDIA Video Search and Summarization Blueprint Claim

  The NVIDIA Blueprint for Video Search and Summarization (VSS) is a set of reference architectures for building vision agents and video analytics applications. It combines accelerated vision microservices, vision language models and LLMs for real-time video intelligence, downstream analytics and agentic search, summarisation and Q&A over stored or streamed video. NVIDIA advises deploying it inside a trusted, isolated network.

• OpenReasoning-Nemotron-32B Claim

  OpenReasoning-Nemotron-32B is an NVIDIA reasoning model derived from Qwen2.5-32B and post-trained for maths, code and science solution generation, evaluated with up to 64K output tokens. It is part of the OpenReasoning-Nemotron family, which also comes in 1.5B, 7B and 14B sizes. The models can run in a heavier mode that starts several parallel generations and combines them with generative solution selection (GenSelect). It is published on Hugging Face under the CC BY 4.0 licence and is ready for commercial and non-commercial research use.

• Parakeet-TDT-0.6B-v3 Claim

  parakeet-tdt-0.6b-v3 is a 600-million-parameter multilingual automatic speech recognition model from NVIDIA, built for high-throughput speech-to-text transcription. It extends parakeet-tdt-0.6b-v2 from English to 25 European languages, including English, French, German, Spanish, Italian, Polish and Ukrainian, and detects the spoken language automatically without prompting. Features include automatic punctuation and capitalisation, word- and segment-level timestamps, and transcription of long audio of up to 24 minutes with full attention or up to three hours with local attention. It uses a FastConformer-TDT architecture, is trained mainly on NVIDIA's Granary corpus and is released under CC BY 4.0 for commercial and non-commercial use.

• RADIOv2.5-L Claim

  NVIDIA sells GPUs, CUDA software and AI enterprise stacks used for training and inference. Public pages cover data-centre, cloud and on-prem AI compute. RADIOv2.5-L is NVIDIA's model API, from NVIDIA Corporation. Published at huggingface.co/nvidia/RADIOv2.5-L. Directory filing: Vision models. Vendor legal name on this product record: NVIDIA Corporation. RADIOv2.5-L is NVIDIA's model API, from NVIDIA Corporation. Published at huggingface.co/nvidia/RADIOv2.5-L. Directory filing: Vision models.

Why complete verification

Completing verification can improve the quality of the public evidence available on this product profile. Useful verified information can deepen the profile and make public questions about the named product more answerable. It does not buy directory ranking, review stars, a higher Assurance Score, a favourable certificate outcome, or a guaranteed Google ranking.

Payment remains independent from rankings and assessment outcomes. Start an assessment · Methodology