The Hao AI Lab research team at the University of California San Diego  — at the forefront of pioneering AI model innovation — recently received an NVIDIA DGX B200 system to elevate their critical work in large language model inference.

Many LLM inference platforms in production today, such as NVIDIA Dynamo, use research concepts that originated in the Hao AI Lab, including DistServe.

With the DGX B200 now fully accessible to the Hao AI Lab and broader UC San Diego community at the School of Computing, Information and Data Sciences’ San Diego Supercomputer Center, the research opportunities are boundless.

“DGX B200 is one of the most powerful AI systems from NVIDIA to date, which means that its performance is among the best in the world,” said Hao Zhang, assistant professor in the Halıcıoğlu Data Science Institute and department of computer science and engineering at UC San Diego. “It enables us to prototype and experiment much faster than using previous-generation hardware.”

Two Hao AI Lab projects the DGX B200 is accelerating are FastVideo and the Lmgame benchmark.

FastVideo focuses on training a family of video generation models to produce a five-second video based on a given text prompt — in just five seconds.

The research phase of FastVideo taps into NVIDIA H200 GPUs in addition to the DGX B200 system.

Lmgame-bench is a benchmarking suite that puts LLMs to the test using popular online games including Tetris and Super Mario Bros. Users can test one model at a time or put two models up against each other to measure their performance.

Other ongoing projects at Hao AI Labs explore new ways to achieve low-latency LLM serving, pushing large language models toward real-time responsiveness.

“Our current research uses the DGX B200 to explore the next frontier of low-latency LLM-serving on the awesome hardware specs the system gives us,” said Junda Chen, a doctoral candidate in computer science at UC San Diego.

Disaggregated inference is a way to ensure large-scale LLM-serving engines can achieve the optimal aggregate system throughput while maintaining acceptably low latency for user requests.

The benefit of disaggregated inference lies in optimizing what DistServe calls “goodput” instead of “throughput” in the LLM-serving engine.

Here’s the difference:

Throughput is measured by the number of tokens per second that the entire system can generate. Higher throughput means lower cost to generate each token to serve the user. For a long time, throughput was the only metric used by LLM-serving engines to measure their performance against one another.

While throughput measures the aggregate performance of the system, it doesn’t directly correlate to the latency that a user perceives. If a user demands lower latency to generate the tokens, the system has to sacrifice throughput.

This natural trade-off between throughput and latency is what led the DistServe team to propose a new metric, “goodput”: the measure of throughput while satisfying the user-specified latency objectives, usually called service-level objectives. In other words, goodput represents the overall health of a system while satisfying user experience.

DistServe shows that goodput is a much better metric for LLM-serving systems, as it factors in both cost and service quality. Goodput leads to optimal efficiency and ideal output from a model.

When a user makes a request in an LLM system, the system takes the user input and generates the first token, known as prefill. Then, the system creates numerous output tokens, one after another, predicting each token’s future behavior based on past requests’ outcomes. This process is known as decode.

Prefill and decode have historically run on the same GPU, but the researchers behind DistServe found that splitting them onto different GPUs maximizes goodput.

“Previously, if you put these two jobs on a GPU, they would compete with each other for resources, which could make it slow from a user perspective,” Chen said. “Now, if I split the jobs onto two different sets of GPUs — one doing prefill, which is compute intensive, and the other doing decode, which is more memory intensive — we can fundamentally eliminate the interference between the two jobs, making both jobs run faster.

This process is called prefill/decode disaggregation, or separating the prefill from decode to get greater goodput.

Increasing goodput and using the disaggregated inference method enables the continuous scaling of workloads without compromising on low-latency or high-quality model responses.

NVIDIA Dynamo — an open-source framework designed to accelerate and scale generative AI models at the highest efficiency levels with the lowest cost — enables scaling disaggregated inference.

In addition to these projects, cross-departmental collaborations, such as in healthcare and biology, are underway at UC San Diego to further optimize an array of research projects using the NVIDIA DGX B200, as researchers continue exploring how AI platforms can accelerate innovation.

Learn more about the NVIDIA DGX B200 system. 

Editor’s note: This blog has been updated to showcase additional recent innovations tapping into the NVIDIA Jetson platform.

Developers, researchers, hobbyists and students can take a byte out of holiday shopping this season as NVIDIA has unwrapped special discounts on the NVIDIA Jetson family of developer kits for edge AI and robotics — available through Sunday, Jan. 11.

Whether tapping into the breakthrough capabilities of Jetson AGX Thor, the versatility of Jetson AGX Orin or the palm-sized power of the NVIDIA Jetson Orin Nano Super Developer Kit, anyone can deck out their bots with greater physical AI performance for lower costs:

Jump to the below sections showcasing how the NVIDIA Jetson platform presents the future of robotics:

Robotics enthusiast Dave Niewinski has built a self‑paddling canoe using the Jetson Orin Nano Super Developer Kit, letting boaters relax and glide on their rides.

The robotic boat integrates two six-axis robotic arms mounted on a lightweight canoe frame, with paddle motion controlled through ROS software and AI algorithms.

The Jetson Orin Nano Super delivers up to 67 INT8 TOPS of AI performance through its NVIDIA Ampere architecture GPU, 32 Tensor Cores and 1,024 CUDA cores, alongside a six-core Arm Cortex‑A78AE CPU and 8GB LPDDR5 memory with 102 GB/s bandwidth.

Its low power envelope of up to 25 watts enables sustained real-time inference and control in mobile, battery-powered applications.

Roboticist Frank Fu built a replica of MIT’s Mini Cheetah robotic quadruped from scratch using the energy-efficient Jetson Orin Nano Super Developer Kit.

This super robot dog can run, trot on uneven surfaces and perform Olympics-worthy backflips.

The secret to its agility lies in the integration of Jetson Orin Nano Super with Fu’s custom quadruped to deliver real-time inference and navigation capabilities.

With GPU-accelerated AI, the robot can recognize objects, follow paths and execute smooth gaits using edge-processed vision and control algorithms. In other words, equipped with the Jetson Orin Nano Super, it can walk like a real dog — and do much more.

Developer Kabilan Kb is advancing warehouse automation with a fully automated forklift, built and tested in the NVIDIA Isaac Sim framework and powered by Jetson Orin Nano Super and Isaac ROS.

The forklift features physics-accurate wheel steering and mast lifting, letting it navigate and work through tasks in a simulated warehouse setting. Virtual cameras and lidar sensors stream data to ROS 2 for mapping, obstacle avoidance and perception testing — preparing control logic for real-world deployment on the NVIDIA Jetson platform.

This workflow enables robotics teams to test and validate navigation, safety and automation capabilities in simulation, dramatically reducing development time before rolling into real warehouses.

In the video below, robot maker Nikodem Bartnik showcases an AI-driven rover built by combining a Raspberry Pi for peripherals with the Jetson Orin Nano Super as the edge AI source.

This little rover runs onboard object-detection and autonomous-navigation algorithms with the Jetson Orin Nano Super delivering fast, low-latency inference.

The robot can read books, navigate mazes and make friends by speaking with humans and fellow robots using ElevenLabs’ AI voice model.

The Jetson Orin Nano Super enables sophisticated robotics workloads in a compact, battery-friendly kit for field exploration. It also lets everything run locally on the rover, without needing to be connected to an external server to operate the large language models for reading and speech.

OptoScale, an AI aquaculture company based in Norway, has integrated the Jetson AGX Orin-powered MX13/23 platform from edge AI solutions provider Aetina to build an underwater AI sensing system that monitors fish health in massive open-sea pens housing thousands of fish.

Mounted inside a submerged camera module, the Jetson AGX Orin processes high-resolution video streams directly at the edge and runs real-time vision models on the device. This allows the system to estimate biomass with exceptional accuracy and deliver continuous inference-based insights even in remote environments with limited connectivity.

The Jetson AGX Orin delivers powerful GPU acceleration and high TOPS AI performance in a compact, energy-efficient module ideal for deployment anywhere.

Las Vegas-based Richtech Robotics is developing Dex, a mobile humanoid robot for factory and warehouse environments designed to handle light-to medium-weight industrial tasks like machine operation, parts sorting, material handling and packaging.

Running on NVIDIA Jetson AGX Thor, Dex combines the mobility of an autonomous wheeled platform with the precision of dual-arm dexterity, allowing it to efficiently navigate environments and pick and place objects. It was trained with a mix of real-world and synthetic data generated from NVIDIA Isaac Sim.

NVIDIA Jetson Thor modules enable real-time reasoning for physical AI, delivering up to 2,070 FP4 teraflops of AI compute and 128GB of memory with power configurable between 40-130 watts.

With discounted pricing for a limited time and a full spectrum of performance options, the NVIDIA Jetson family gives anyone the tools to design, build and deploy the next generation of intelligent machines. It’s the ideal gift for robot lovers — and the robots in their lives.

Shop NVIDIA Jetson now.

A look under the hood of virtually any frontier model today will reveal a mixture-of-experts (MoE) model architecture that mimics the efficiency of the human brain.

Just as the brain activates specific regions based on the task, MoE models divide work among specialized “experts,” activating only the relevant ones for every AI token. This results in faster, more efficient token generation without a proportional increase in compute.

The industry has already recognized this advantage. On the independent Artificial Analysis (AA) leaderboard, the top 10 most intelligent open-source models use an MoE architecture, including DeepSeek AI’s DeepSeek-R1, Moonshot AI’s Kimi K2 Thinking, OpenAI’s gpt-oss-120B and Mistral AI’s Mistral Large 3.

However, scaling MoE models in production while delivering high performance is notoriously difficult. The extreme codesign of NVIDIA GB200 NVL72 systems combines hardware and software optimizations for maximum performance and efficiency, making it practical and straightforward to scale MoE models.

The Kimi K2 Thinking MoE model — ranked as the most intelligent open-source model on the AA leaderboard — sees a 10x performance leap on the NVIDIA GB200 NVL72 rack-scale system compared with NVIDIA HGX H200. Building on the performance delivered for the DeepSeek-R1 and Mistral Large 3 MoE models, this breakthrough underscores how MoE is becoming the architecture of choice for frontier models — and why NVIDIA’s full-stack inference platform is the key to unlocking its full potential.

Until recently, the industry standard for building smarter AI was simply building bigger, dense models that use all of their model parameters — often hundreds of billions for today’s most capable models — to generate every token. While powerful, this approach requires immense computing power and energy, making it challenging to scale.

Much like the human brain relies on specific regions to handle different cognitive tasks — whether processing language, recognizing objects or solving a math problem — MoE models comprise several specialized “experts.” For any given token, only the most relevant ones are activated by a router. This design means that even though the overall model may contain hundreds of billions of parameters, generating a token involves using only a small subset — often just tens of billions.

By selectively engaging only the experts that matter most, MoE models achieve higher intelligence and adaptability without a matching rise in computational cost. This makes them the foundation for efficient AI systems optimized for performance per dollar and per watt — generating significantly more intelligence for every unit of energy and capital invested.

Given these advantages, it is no surprise that MoE has rapidly become the architecture of choice for frontier models, adopted by over 60% of open-source AI model releases this year. Since early 2023, it’s enabled a nearly 70x increase in model intelligence — pushing the limits of AI capability.

“Our pioneering work with OSS mixture-of-experts architecture, starting with Mixtral 8x7B two years ago, ensures advanced intelligence is both accessible and sustainable for a broad range of applications,” said Guillaume Lample, cofounder and chief scientist at Mistral AI. “Mistral Large 3’s MoE architecture enables us to scale AI systems to greater performance and efficiency while dramatically lowering energy and compute demands.”

Frontier MoE models are simply too large and complex to be deployed on a single GPU. To run them, experts must be distributed across multiple GPUs, a technique called expert parallelism. Even on powerful platforms such as the NVIDIA H200, deploying MoE models involves bottlenecks such as:

The solution: extreme codesign.

NVIDIA GB200 NVL72 is a rack-scale system with 72 NVIDIA Blackwell GPUs working together as if they were one, delivering 1.4 exaflops of AI performance and 30TB of fast shared memory. The 72 GPUs are connected using NVLink Switch into a single, massive NVLink interconnect fabric, which allows every GPU to communicate with each other with 130 TB/s of NVLink connectivity.

MoE models can tap into this design to scale expert parallelism far beyond previous limits — distributing the experts across a much larger set of up to 72 GPUs.

This architectural approach directly resolves MoE scaling bottlenecks by:

Other full-stack optimizations also play a key role in unlocking high inference performance for MoE models. The NVIDIA Dynamo framework orchestrates disaggregated serving by assigning prefill and decode tasks to different GPUs, allowing decode to run with large expert parallelism, while prefill uses parallelism techniques better suited to its workload. The NVFP4 format helps maintain accuracy while further boosting performance and efficiency.

Open-source inference frameworks such as NVIDIA TensorRT-LLM, SGLang and vLLM support these optimizations for MoE models. SGLang, in particular, has played a significant role in advancing large-scale MoE on GB200 NVL72, helping validate and mature many of the techniques used today.

To bring this performance to enterprises worldwide, the GB200 NVL72 is being deployed by  major cloud service providers and NVIDIA Cloud Partners including Amazon Web Services, Core42, CoreWeave, Crusoe, Google Cloud, Lambda, Microsoft Azure, Nebius, Nscale, Oracle Cloud Infrastructure, Together AI and others.

“At CoreWeave, our customers are leveraging our platform to put mixture-of-experts models into production as they build agentic workflows,” said Peter Salanki, cofounder and chief technology officer at CoreWeave. “By working closely with NVIDIA, we are able to deliver a tightly integrated platform that brings MoE performance, scalability and reliability together in one place. You can only do that on a cloud purpose-built for AI.”

Customers such as DeepL are using Blackwell NVL72 rack-scale design to build and deploy their next-generation AI models.

“DeepL is leveraging NVIDIA GB200 hardware to train mixture-of-experts models, advancing its model architecture to improve efficiency during training and inference, setting new benchmarks for performance in AI,” said Paul Busch, research team lead at DeepL.

NVIDIA GB200 NVL72 efficiently scales complex MoE models and delivers a 10x leap in performance per watt. This performance leap isn’t just a benchmark; it enables 10x the token revenue, transforming the economics of AI at scale in power- and cost-constrained data centers.

At NVIDIA GTC Washington, D.C., NVIDIA founder and CEO Jensen Huang highlighted how GB200 NVL72 delivers 10x the performance of NVIDIA Hopper for DeepSeek-R1, and this performance extends to other DeepSeek variants as well.

“With GB200 NVL72 and Together AI’s custom optimizations, we are exceeding customer expectations for large-scale inference workloads for MoE models like DeepSeek-V3,” said Vipul Ved Prakash, cofounder and CEO of Together AI. “The performance gains come from NVIDIA’s full-stack optimizations coupled with Together AI Inference breakthroughs across kernels, runtime engine and speculative decoding.”

This performance advantage is evident across other frontier models.

Kimi K2 Thinking, the most intelligent open source model, serves as another proof point, achieving 10x better generational performance when deployed on GB200 NVL72.

Fireworks AI has currently deployed Kimi K2 on the NVIDIA B200 platform to achieve the highest performance on the Artificial Analysis leaderboard.

“NVIDIA GB200 NVL72 rack-scale design makes MoE model serving dramatically more efficient,” said Lin Qiao, cofounder and CEO of Fireworks AI. “Looking ahead, NVL72 has the potential to transform how we serve massive MoE models, delivering major performance improvements over the Hopper platform and setting a new bar for frontier model speed and efficiency.”

The new Kinnie model also achieved a 10x performance gain on the GB200 NVL72 compared with the prior-generation H200. This generational gain translates into better user experience, lower per-token cost and higher energy efficiency for this new MoE model.

The NVIDIA GB200 NVL72 rack-scale system is designed to deliver strong performance beyond MoE models.

The reason becomes clear when taking a look at where AI is heading: the newest generation of multimodal AI models have specialized components for language, vision, audio and other modalities, activating only the ones relevant to the task at hand.

In agentic systems, different “agents” specialize in planning, perception, reasoning, tool use or search, and an orchestrator coordinates them to deliver a single outcome. In both cases, the core pattern mirrors MoE: route each part of the problem to the most relevant experts, then coordinate their outputs to produce the final outcome.

Extending this principle to production environments where multiple applications and agents serve multiple users unlocks new levels of efficiency. Instead of duplicating massive AI models for every agent or application, this approach can enable a shared pool of experts accessible to all, with each request routed to the right expert.

Mixture of experts is a powerful architecture moving the industry toward a future where massive capability, efficiency and scale coexist. The GB200 NVL72 unlocks this potential today, and NVIDIA’s roadmap with the NVIDIA Vera Rubin architecture will continue to expand the horizons of frontier models.

Learn more about how GB200 NVL72 scales complex MoE models in this technical deep dive.

This post is part of Think SMART, a series focused on how leading AI service providers, developers and enterprises can boost their inference performance and return on investment with the latest advancements from NVIDIA’s full-stack inference platform.

Researchers worldwide rely on open-source technologies as the foundation of their work. To equip the community with the latest advancements in digital and physical AI, NVIDIA is further expanding its collection of open AI models, datasets and tools — with potential applications in virtually every research field.

At NeurIPS, one of the world’s top AI conferences, NVIDIA is unveiling open physical AI models and tools to support research, including Alpamayo-R1, the world’s first industry-scale open reasoning vision language action (VLA) model for autonomous driving. In digital AI, NVIDIA is releasing new models and datasets for speech and AI safety.

NVIDIA researchers are presenting over 70 papers, talks and workshops at the conference, sharing innovative projects that span AI reasoning, medical research, autonomous vehicle (AV) development and more.

These initiatives deepen NVIDIA’s commitment to open source — an effort recognized by a new Openness Index from Artificial Analysis, an independent organization that benchmarks AI. The Artificial Analysis Open Index rates the NVIDIA Nemotron family of open technologies for frontier AI development among the most open in the AI ecosystem based on the permissibility of the model licenses, data transparency and availability of technical details.

NVIDIA DRIVE Alpamayo-R1 (AR1), the world’s first open reasoning VLA model for AV research, integrates chain-of-thought AI reasoning with path planning — a component critical for advancing AV safety in complex road scenarios and enabling level 4 autonomy.

While previous iterations of self-driving models struggled with nuanced situations — a pedestrian-heavy intersection, an upcoming lane closure or a double-parked vehicle in a bike lane — reasoning gives autonomous vehicles the common sense to drive more like humans do.

AR1 accomplishes this by breaking down a scenario and reasoning through each step. It considers all possible trajectories, then uses contextual data to choose the best route.

For example, by tapping into the chain-of-thought reasoning enabled by AR1, an AV driving in a pedestrian-heavy area next to a bike lane could take in data from its path, incorporate reasoning traces — explanations on why it took certain actions — and use that information to plan its future trajectory, such as moving away from the bike lane or stopping for potential jaywalkers.

AR1’s open foundation, based on NVIDIA Cosmos Reason, lets researchers customize the model for their own non-commercial use cases, whether for benchmarking or building experimental AV applications.

For post-training AR1, reinforcement learning has proven especially effective — researchers observed a significant improvement in reasoning capabilities with AR1 compared with the pretrained model.

NVIDIA DRIVE Alpamayo-R1 is now available on GitHub and Hugging Face, and a subset of the data used to train and evaluate the model is available in the NVIDIA Physical AI Open Datasets. NVIDIA has also released the open-source AlpaSim framework to evaluate AR1.

Learn more about reasoning VLA models for autonomous driving.

Developers can learn how to use and post-train Cosmos-based models using step-by-step recipes, quick-start inference examples and advanced post-training workflows now available in the Cosmos Cookbook. It’s a comprehensive guide for physical AI developers that covers every step in AI development, including data curation, synthetic data generation and model evaluation.

There are virtually limitless possibilities for Cosmos-based applications. The latest examples from NVIDIA include:

Policy models can be trained in NVIDIA Isaac Lab and Isaac Sim , and data generated from the policy models can then be used to post-train NVIDIA GR00T N models for robotics.

NVIDIA ecosystem partners are developing their latest technologies with Cosmos WFMs.

AV developer Voxel51 is contributing model recipes to the Cosmos Cookbook. Physical AI developers 1X, Figure AI, Foretellix, Gatik, Oxa, PlusAI and X-Humanoid are using WFMs for their latest physical AI applications. And researchers at ETH Zurich are presenting a NeurIPS paper that highlights using Cosmos models for realistic and cohesive 3D scene creation.

NVIDIA Nemotron Additions Bolster the Digital AI Developer Toolkit

NVIDIA is also releasing new multi-speaker speech AI models, a new model with reasoning capabilities and datasets for AI safety, as well as open tools to generate high-quality synthetic datasets for reinforcement learning and domain-specific model customization. These tools include:

NVIDIA ecosystem partners using NVIDIA Nemotron and NeMo tools to build secure, specialized agentic AI include CrowdStrike, Palantir and ServiceNow.

NeurIPS attendees can explore these innovations at the Nemotron Summit, taking place today, from 4-8 p.m. PT, with an opening address by Bryan Catanzaro, vice president of applied deep learning research at NVIDIA.

Of the dozens of NVIDIA-authored research papers at NeurIPS, here are a few highlights advancing language models:

View the full list of events at NeurIPS, running through Sunday, Dec. 7, in San Diego.   

See notice regarding software product information.

Editor’s note: This post is part of the AI On blog series, which explores the latest techniques and real-world applications of agentic AI, chatbots and copilots. The series also highlights the NVIDIA software and hardware powering advanced AI agents, which form the foundation of AI query engines that gather insights and perform tasks to transform everyday experiences and reshape industries.

Today’s computer vision systems excel at identifying what happens in physical spaces and processes, but lack the abilities to explain the details of a scene and why they matter, as well as reason about what might happen next.

Agentic intelligence powered by vision language models (VLMs) can help bridge this gap, giving teams quick, easy access to key insights and analyses that connect text descriptors with spatial-temporal information and billions of visual data points captured by their systems every day.

Three approaches organizations can use to boost their legacy computer vision systems with agentic intelligence are to:

Traditional convolutional neural network (CNN)-powered video search tools are constrained by limited training, context and semantics, making gleaning insights manual, tedious and time-consuming. CNNs are tuned to perform specific visual tasks, like spotting an anomaly, and lack the multimodal ability to translate what they see into text.

Businesses can embed VLMs directly into their existing applications to generate highly detailed captions of images and videos. These captions turn unstructured content into rich, searchable metadata, enabling visual search that’s far more flexible — not constrained by file names or basic tags.

For example, automated vehicle-inspection system UVeye processes over 700 million high-resolution images each month to build one of the world’s largest vehicle and component datasets. By applying VLMs, UVeye converts this visual data into structured condition reports, detecting subtle defects, modifications or foreign objects with exceptional accuracy and reliability for search.

VLM-powered visual understanding adds essential context, ensuring transparent, consistent insights for compliance, safety and quality control. UVeye detects 96% of defects compared with 24% using manual methods, enabling early intervention to reduce downtime and control maintenance costs.

Relo Metrics, a provider of AI-powered sports marketing measurement, helps brands quantify the value of their media investments and optimize their spending. By combining VLMs with computer vision, Relo Metrics moves beyond basic logo detection to capture context — like a courtside banner shown during a game-winning shot — and translate it into real-time monetary value.

This contextual-insight capability highlights when and how logos appear, especially in high-impact moments, giving marketers a clearer view of return on investment and ways to optimize strategy. For example, Stanley Black & Decker, including its Dewalt brand, previously relied on end-of-season reports to evaluate sponsor asset performance, limiting timely decision-making. Using Relo Metrics for real-time insights, Stanley Black & Decker adjusted signage positioning and saved $1.3 million in potentially lost sponsor media value.

CNN-based computer vision systems often generate binary detection alerts such as yes or no, and true or false. Without the reasoning power of VLMs, that can mean false positives and missed details — leading to costly mistakes in safety and security, as well as lost business intelligence.Rather than replacing these CNN-based computer vision systems entirely, VLMs can easily augment these systems as an intelligent add-on. With a VLM layered on top of CNN-based computer vision systems, detection alerts are not only flagged but reviewed with contextual understanding — explaining where, how and why the incident occurred.

For smarter city traffic management, Linker Vision uses VLMs to verify critical city alerts, such as traffic accidents, flooding, or falling poles and trees from storms. This reduces false positives and adds vital context to each event to improve real-time municipal response.

Linker Vision’s architecture for agentic AI involves automating event analysis from over 50,000 diverse smart city camera streams to enable cross-department remediation — coordinating actions across teams like traffic control, utilities and first responders when incidents occur. The ability to query across all camera streams simultaneously enables systems to quickly and automatically turn observations into insights and trigger recommendations for next best actions.

Agentic AI systems can process, reason and answer complex queries across video streams and modalities — such as audio, text, video and sensor data. This is possible by combining VLMs with reasoning models, large language models (LLMs), retrieval-augmented generation (RAG), computer vision and speech transcription.

Basic integration of a VLM into an existing computer vision pipeline is helpful in verifying short video clips of key moments. However this approach is limited by how many visual tokens a single model can process at once, resulting in surface-level answers without context over longer time periods and external knowledge.

In contrast, whole architectures built on agentic AI enable scalable, accurate processing of lengthy and multichannel video archives. This leads to deeper, more accurate and more reliable insights that go beyond surface-level understanding. Agentic systems can be used for root-cause analysis or analysis of long inspection videos to generate reports with timestamped insights.

Levatas develops visual-inspection solutions that use mobile robots and autonomous systems to enhance safety, reliability and performance of critical infrastructure assets such as electric utility substations, fuel terminals, rail yards and logistics hubs. Using VLMs, Levatas built a video analytics AI agent to automatically review inspection footage and draft detailed inspection reports, dramatically accelerating a traditionally manual and slow process.

For customers like American Electric Power (AEP), Levatas AI integrates with Skydio X10 devices to streamline inspection of electric infrastructure. Levatas enables AEP to autonomously inspect power poles, identify thermal issues and detect equipment damage. Alerts are sent instantly to the AEP team upon issue detection, enabling swift response and resolution, and ensuring reliable, clean and affordable energy delivery.

AI gaming highlight tools like Eklipse use VLM-powered agents to enrich livestreams of video games with captions and index metadata for rapid querying, summarization and creation of polished highlight reels in minutes — 10x faster than legacy solutions — leading to improved content consumption experiences.

For advanced search and reasoning, developers can use multimodal VLMs such as NVCLIP, NVIDIA Cosmos Reason and Nemotron Nano V2 to build metadata-rich indexes for search.

To integrate VLMs into computer vision applications, developers can use the event reviewer feature in the NVIDIA Blueprint for video search and summarization (VSS), part of the NVIDIA Metropolis platform.

For more complex queries and summarization tasks, the VSS blueprint can be customized so developers can build AI agents that access VLMs directly or use VLMs in conjunction with LLMs, RAG and computer vision models. This enables smarter operations, richer video analytics and real-time process compliance that scale with organizational needs.

Learn more about NVIDIA-powered agentic video analytics.

Stay up to date by subscribing to NVIDIA’s vision AI newsletter, joining the community and following NVIDIA AI on LinkedIn, Instagram, X and Facebook.  

Explore the VLM tech blogs, and self-paced video tutorials and livestreams.

Two out of every three people are likely to be living in cities or other urban centers by 2050, according to the United Nations, meaning about 2.5 billion people could be added to urban areas by the middle of the century. This highlights an urgent need for more sustainable urban planning and public services.

The smart traffic management market alone is projected to reach $20 billion by 2027 as cities deploy AI to accommodate growing population density and tourism-driven congestion.

Improved traffic management is just one way AI is enhancing public spaces, transit and city processes — with many physical AI technologies being showcased at the Smart City Expo World Congress (SCEWC) taking place this week in Barcelona.

At the event, NVIDIA’s expanding ecosystem of physical AI partners will be on full display —  including simulation and mapping companies, software vendors, manufacturers, systems integrators, and cloud and edge providers — all exhibiting their latest smart spaces applications.

The NVIDIA Blueprint for smart city AI — announced in June at NVIDIA GTC Paris — combines digital twins with NVIDIA Omniverse libraries, synthetic data generation, AI model training with vision language models (VLMs) and video analytics AI agents using the NVIDIA Blueprint for video search and summarization (VSS).

These work together in a single, streamlined workflow, enabling rapid simulation of real-world urban conditions, vast sensor integration and real-time vision AI.

Updates to the blueprint include new NVIDIA Cosmos world foundation models and VLMs to generate photorealistic synthetic data and provide physical reasoning capabilities, as well as the latest version of the VSS blueprint  — part of the NVIDIA Metropolis platform for training and building vision AI agents and applications.

Access step-by-step intelligent traffic system workflows and technical recipes in new cookbooks for NVIDIA Cosmos Predict, NVIDIA Cosmos Transfer and NVIDIA Cosmos Reason. 

Learn more about how five leading companies are using these NVIDIA technologies to bring secure, scalable physical AI to more cities, including Dublin, Ireland; Ho Chi Minh City, Vietnam; and Raleigh, North Carolina.

Urban populations are soaring, infrastructure is aging, and challenges like traffic congestion, climate resilience and real-time safety demand new levels of responsiveness and intelligence.

But integrating these video- and data-intensive systems isn’t just a matter of upgrading old infrastructure. Smart city deployments require a seamless fusion of next-generation AI, cloud and edge computing, ultrafast networking and open data platforms. A global ecosystem of partners, such as the ones highlighted below, must come together to solve these technological and real-world challenges for city-scale AI solutions.

Mapping and spatial analytics company Esri is working with NVIDIA and Microsoft to build an AI agent that ingests, analyzes and visualizes massive amounts of camera data to generate real-time alerts and insights in an interactive geospatial map for the City of Raleigh in North Carolina.

An interactive geospatial map of Raleigh with a video analytics AI agent processing live sensor data. Images courtesy of Esri and the City of Raleigh.

This will enable city operators to quickly respond and notify key departments to efficiently resolve issues, optimize traffic flows and improve infrastructure design, boosting overall urban mobility. Tapping into components of the NVIDIA Blueprint for smart city AI will enable the city to discover, aggregate and manipulate data to help address key resident concerns, automate streetlight timing to reduce road delays and decrease carbon emissions from cars sitting in traffic.

Milestone Systems is now introducing generative AI to its XProtect video management platform — allowing users to easily extract valuable analyses and insights from their video streams and libraries, review alerts and automatically generate reports. This functionality is powered by NVIDIA Cosmos Reason VLMs that Milestone post-trained with 75,000 hours of compliant traffic video, for each VLM, from either Europe or the U.S.

Milestone is also offering these specialized VLMs-as-a-service to developers to build innovative applications. General availability is expected later this year.

XProtect customers can easily add these new, GDPR-compliant features and add an additional layer of intelligence by contextualizing video, scheduling insights and providing on-demand reporting and automatic review of alerts. These features could reduce up to 30% of operator alarm fatigue by automating video review and filtering out false alarms. Cities like Dubuque, Iowa, and Genoa, Italy, are planning to try these new capabilities, adopting advanced video intelligence solutions to enhance traffic management.

Linker Vision, a pioneer in deploying the end-to-end NVIDIA Blueprint for smart city AI, is bringing physical AI to Ho Chi Minh City and Danang in Vietnam.

Building on its success in Kaohsiung City, Taiwan, where its vision AI platform cut incident response times by up to 80%, Linker Vision, tapping into AVES Reality’s simulation-ready 3D digital twins using NVIDIA Omniverse, is now harnessing the blueprint to develop city-scale AI solutions that can simulate and monitor traffic and construction activities, helping ensure projects stay on schedule while improving urban mobility and operational efficiency.

Ireland’s Smart Dublin organization addresses urban challenges with physical AI and is tapping into Cesium, a 3D geospatial platform from Bentley Systems, and NVIDIA Omniverse to enable real-time data visualization and analysis for micromobility — like walking, cycling and using scooters — across the city.

In addition, AI transportation technology company VivaCity is using the NVIDIA Jetson and Metropolis platforms on its AI-powered computer vision sensors, providing Dublin with highly accurate, multimodal data and insights on cyclists, motor vehicles and pedestrians. This data is used to understand traffic patterns and road user behavior, as well as to identify dangerous sites using VivaCity’s award-winning solution across the Irish capital’s roadways, improving road safety and traffic flow.

Video courtesy of Smart Dublin and Bentley Systems.

Smart Dublin is also working with partners to integrate AI agents into a digital twin to gain insights and alert city operators faster. As a first step, Smart Dublin integrated VivaCity data into a Cesium-powered digital twin to visualize and understand bike route utilization. Adding rainfall data, the city saw that poor weather had a negligible impact on the number of cyclists on the road. Future plans include adding AI features like natural language search.

In addition to the Smart Dublin project, Bentley’s AI technology Blyncsy uses NVIDIA Cosmos and Metropolis to generate synthetic data for analyzing road conditions and improving maintenance.

Deloitte is applying AI to automate street inspections across thousands of crosswalks for city transportation departments, aiming to protect vulnerable roadway users such as pedestrians and bikers.

NVIDIA Cosmos Transfer simulates real-world conditions and generates videos for different scenarios. Video courtesy of Deloitte.

With Cosmos Predict, Deloitte converts static intersection images into photorealistic, physically accurate videos. Then, Cosmos Transfer generates videos across a mixture of real-world scenarios, such as fog, rain, snow and low light — bringing each environment to life. Finally, Cosmos Reason helps evaluate these videos and identify potential areas of improvement, making the process insightful and dynamic.

The NVIDIA hardware and other ecosystem partners making such smart city deployments possible include AAEON, Advantech, Aetina, Dell Technologies, HPE, OpenZeka and YUAN High Technologies.

Many of these partners will exhibit physical AI demos running on NVIDIA RTX PRO Servers, NVIDIA DGX Spark — the world’s smallest AI supercomputer — as well as NVIDIA Jetson Thor modules on the show floor.

Plus, see Akila’s physical AI technologies in the Monaco-Monte-Carlo train stations, as well as its digital twin solutions for the University Mohammed VI Polytechnic campus in Morocco.

And K2K will be demonstrating its real-time analytics solution that uses NVIDIA Cosmos Reason and the VSS blueprint to optimize roadway safety in addition to traffic and waste management.

See city-scale AI and urban innovation in action by joining NVIDIA and its partners at SCEWC and sign up to be notified when the NVIDIA Blueprint for smart city AI becomes available.

Stay up to date by following NVIDIA AI on LinkedIn, Instagram, X and Facebook.

Editor’s note: This post is part of Into the Omniverse, a series focused on how developers, 3D practitioners and enterprises can transform their workflows using the latest advances in OpenUSD and NVIDIA Omniverse.

Physical AI models — which power robots, autonomous vehicles and other intelligent machines — must be safe, generalized for dynamic scenarios and capable of perceiving, reasoning and operating in real time. Unlike large language models that can be trained on massive datasets from the internet, physical AI models must learn from data grounded in the real world.

However, collecting sufficient data that covers this wide variety of scenarios in the real world is incredibly difficult and, in some cases, dangerous. Physically based synthetic data generation offers a key way to address this gap.

NVIDIA recently released updates to NVIDIA Cosmos open world foundation models (WFMs) to accelerate data generation for testing and validating physical AI models. Using NVIDIA Omniverse libraries and Cosmos, developers can generate physically based synthetic data at incredible scale.

Cosmos Predict 2.5 now unifies three separate models — Text2World, Image2World and Video2World — into a single lightweight architecture that generates consistent, controllable multicamera video worlds from a single image, video or prompt.

Cosmos Transfer 2.5 enables high-fidelity, spatially controlled world-to-world style transfer to amplify data variation. Developers can add new weather, lighting and terrain conditions to their simulated environments across multiple cameras. Cosmos Transfer 2.5 is 3.5x smaller than its predecessor, delivering faster performance with improved prompt alignment and physics accuracy.

These WFMs can be integrated into synthetic data pipelines running in the NVIDIA Isaac Sim open-source robotics simulation framework, built on the NVIDIA Omniverse platform, to generate photorealistic videos that reduce the simulation-to-real gap. Developers can reference a four-part pipeline for synthetic data generation:

Leading robotics and AI companies are already using these technologies to accelerate physical AI development.

Skild AI, which builds general-purpose robot brains, is using Cosmos Transfer to augment existing data with new variations for testing and validating robotics policies trained in NVIDIA Isaac Lab.

Skild AI uses Isaac Lab to create scalable simulation environments where its robots can train across embodiments and applications. By combining Isaac Lab robotics simulation capabilities with Cosmos’ synthetic data generation, Skild AI can train robot brains across diverse conditions without the time and cost constraints of real-world data collection.

Serve Robotics uses synthetic data generated from thousands of simulated scenarios in NVIDIA Isaac Sim. The synthetic data is then used in conjunction with real data to train physical AI models. The company has built one of the largest autonomous robot fleets operating in public spaces and has completed over 100,000 last-mile meal deliveries across urban areas. Serve’s robots collect 1 million miles of data monthly, including nearly 170 billion image-lidar samples, which are used in simulation to further improve robot models.

Learn more about how Serve Robotics uses Isaac Sim to accelerate development, testing and deployment of its sidewalk delivery robots by watching the below livestream.

Beyond bringing people meals, Serve recently used its robots to deliver compute power — dropping off brand-new NVIDIA DGX Spark personal AI supercomputers to Refik Anadol, Will.I.AM and Ollama. With 1 petaflop of AI performance, DGX Spark offers developers desktop capabilities for workflows from AI model prototyping and model fine-tuning to inference and robotics development.

Autonomous drone delivery company Zipline also participated in the DGX Spark drop, with Chief Hardware Officer Jo Mardall receiving a DGX Spark by drone at the company’s headquarters and testing facility in Half Moon Bay, California. Zipline uses the NVIDIA Jetson edge AI and robotics platform for its drone delivery systems.

Lightwheel, a simulation-first robotics solution provider, is helping companies bridge the simulation-to-real gap with SimReady assets and large-scale synthetic datasets. With high-quality synthetic data and simulation environments built on OpenUSD, Lightwheel’s approach helps ensure robots trained in simulation perform effectively in real-world scenarios, from factory floors to homes.

Data scientist and Omniverse community member Santiago Villa is using synthetic data with Omniverse libraries and Blender software to improve mining operations by identifying large boulders that halt operations.

Undetected boulders entering crushers can cause delays of seven minutes or more per incident, costing mines up to $650,000 annually in lost production. Using Omniverse to generate thousands of automatically annotated synthetic images across varied lighting and weather conditions dramatically reduces training costs while enabling mining companies to improve boulder detection systems and avoid equipment downtime.

FS Studio partnered with a global logistics leader to improve AI-driven package detection by creating thousands of photorealistic package variations in different lighting conditions using Omniverse libraries like Replicator. The synthetic dataset dramatically improved object detection accuracy and reduced false positives, delivering measurable gains in throughput speed and system performance across the customer’s logistics network.

Robots for Humanity built a full simulation environment in Isaac Sim for an oil and gas client using Omniverse libraries to generate synthetic data, including depth, segmentation and RGB images, while collecting joint and motion data from the Unitree G1 robot through teleoperation.

Omniverse Ambassador Scott Dempsey is developing a synthetic data generation synthesizer that builds various cables from real-world manufacturer specifications, using Isaac Sim to generate synthetic data augmented with Cosmos Transfer to create photorealistic training datasets for applications that detect and handle cables.

Learn more about OpenUSD, Cosmos and synthetic data for physical AI by exploring these resources:

Stay up to date by subscribing to NVIDIA Omniverse news, joining the Omniverse community and following Omniverse on Discord, Instagram, LinkedIn, Threads, X and YouTube. 

Explore the Alliance for OpenUSD forum and the AOUSD website.

As Open Source AI Week comes to a close, we’re celebrating the innovation, collaboration and community driving open-source AI forward. Catch up on the highlights and stay tuned for more announcements coming next week at NVIDIA GTC Washington, D.C.

From the stages of the PyTorch Conference to workshops across Open Source AI Week, this week spotlighted the creativity and progress defining the future of open AI.

Here are some highlights from the event:

Honoring open-source contributions: Jonathan Dekhtiar, senior deep learning framework engineer at NVIDIA, received the PyTorch Contributor Award for his key role in designing the release mechanisms and packaging solutions for Python software and libraries that enable GPU-accelerated computing.

CEO of Modular visits the NVIDIA booth: Chris Lattner, CEO of Modular and founder and chief architect of the open-source LLVM Compiler Infrastructure project, picks up the NVIDIA DGX Spark.

Seven questions with founding researcher at fast.ai: Jeremy Howard, founding researcher at fast.ai and advocate for accessible deep learning, shares his insights on the future of open-source AI.

In his keynote at the PyTorch Conference, Howard also highlighted the growing strength of open-source communities, recognizing NVIDIA for its leadership in advancing openly available, high-performing AI models.

“The one company, actually, that has stood out, head and shoulders above the others, and that is two,” he said. “One is Meta…the creators of PyTorch. The other is NVIDIA, who, just in recent months, has created some of the world’s best models — and they are open source, and they are openly licensed.”

Open-source innovation is accelerating. NVIDIA and the vLLM team are partnering to add vLLM upstream support for NVIDIA Nemotron models, transforming open large language model (LLM) serving with lightning-fast performance, efficient scaling and simplified deployment across NVIDIA GPUs.

vLLM’s optimized inference engine empowers developers to run Nemotron models like the new Nemotron Nano 2 — a highly efficient small language reasoning model with a hybrid Transformer-Mamba architecture and a configurable thinking budget.

Learn more about how vLLM is accelerating open model innovation.

NVIDIA is making eight NVIDIA Nemotron RAG models openly available on Hugging Face, expanding access beyond research to include the full suite of commercial models.

This release gives developers a wider range of tools to build retrieval-augmented generation (RAG) systems, improve search and ranking accuracy, and extract structured data from complex documents.

The newly released models include Llama-Embed-Nemotron-8B, which provides multilingual text embeddings built on Llama 3.1, and Omni-Embed-Nemotron-3B, which supports cross-modal retrieval for text, images, audio and video.

Developers can also access six production-grade models for text embedding, reranking and PDF data extraction — key components for real-world retrieval and document intelligence applications.

With these open-source models, developers, researchers and organizations can more easily integrate and experiment with RAG-based systems.

Developers can get started with Nemotron RAG on Hugging Face.

NVIDIA is expanding access to high-quality open datasets that help developers overcome the challenges of large-scale data collection and focus on building advanced AI systems.

The latest release includes a collection of Nemotron-Personas datasets for Sovereign AI. Each dataset is fully synthetic and grounded in real-world demographic, geographic and cultural data — with no personally identifiable information. The growing collection, which features personas from the U.S., Japan and India, enables model builders to design AI agents and systems that reflect the linguistic, social and contextual nuance of the nations they serve.

NVIDIA earlier this year released the NVIDIA Physical AI Open Datasets on Hugging Face, featuring more than 7 million robotics trajectories and 1,000 OpenUSD SimReady assets. Downloaded more than 6 million times, the datasets combines real‑world and synthetic data from the NVIDIA Cosmos,  Isaac, DRIVE and Metropolis platforms to kickstart physical AI development.

At the PyTorch Conference’s Startup Showcase, 11 startups — including members from the NVIDIA Inception program — are sharing their work developing practical AI applications and connecting with investors, potential customers and peers.

Runhouse, an AI infrastructure startup optimizing model deployment and orchestration, was crowned the 2025 PyTorch Startup Showcase Award Winner. The Community Choice Award was presented to CuraVoice, with CEO Sakhi Patel, CTO Shrey Modi, and advisor Rahul Vishwakarma accepting the award on behalf of the team.

CuraVoice provides an AI-powered voice simulation platform — powered by NVIDIA Riva for speech recognition and text-to-speech, and NVIDIA NeMo for conversational AI models — for healthcare students and professionals, offering interactive exercises and adaptive feedback to improve patient communication skills.

In addition to CuraVoice, other Inception members, including Backfield AI, Graphsignal, Okahu AI, Snapshot AI and XOR, were featured participants in the Startup Showcase.

Snapshot AI delivers actionable, real-time insights to engineering teams using recursive RAG, transformers and multimodal AI. The company’s platform taps into the NVIDIA CUDA Toolkit to deliver high-performance analysis and rapid insights at scale.

XOR is a cybersecurity startup offering AI agents that automatically fix vulnerabilities in the supply chain of other AIs. The company helps enterprises eliminate vulnerabilities while complying with regulatory requirements. XOR’s agentic technology uses NVIDIA cuVS vector search for indexing, real-time retrieval and code analysis. The company also uses GPU-based machine learning to train models to detect hidden backdoor patterns and prioritize of high-value security outcomes.

Attendees of Open Source AI Week are getting a peek at the latest advancements and creative projects that are shaping the future of open technology.

Here’s a look at what’s happening onsite:

The world’s smallest AI supercomputer: NVIDIA DGX Spark represents the cutting edge of AI computing hardware for enterprise and research applications.​

Humanoids and robot dogs, up close: Unitree robots are on display, captivating attendees with advanced mobility powered by the latest robotics technology.

Why open source is important: Learn how it can empower developers to build stronger communities, iterate on features, and seamlessly integrate the best of open source AI.

A study from the Center for Security and Emerging Technology (CSET) published today shows how access to open model weights unlocks more opportunities for experimentation, customization and collaboration across the global research community.​

The report outlines seven high-impact research use cases where open models are making a difference — including fine-tuning, continued pretraining, model compression and interpretability.

With access to weights, developers can adapt models for new domains, explore new architectures and extend functionality to meet their specific needs. This also supports trust and reproducibility. When teams can run experiments on their own hardware, share updates and revisit earlier versions, they gain control and confidence in their results.

Additionally, the study found that nearly all open model users share their data, weights and code, building a fast-growing culture of collaboration. This open exchange of tools and knowledge strengthens partnerships between academia, startups and enterprises, facilitating innovation.

NVIDIA is committed to empowering the research community through the NVIDIA Nemotron family of open models — featuring not just open weights, but also pretraining and post-training datasets, detailed training recipes, and research papers that share the latest breakthroughs.

Read the full CSET study to learn how open models are helping the AI community move forward.

At the PyTorch Conference, Jim Fan, director of robotics and distinguished research scientist at NVIDIA, discussed the Physical Turing Test — a way of measuring the performance of intelligent machines in the physical world.

With conversational AI now capable of fluent, lifelike communication, Fan noted that the next challenge is enabling machines to act with similar naturalism. The Physical Turing Test asks: can an intelligent machine perform a real-world task so fluidly that a human cannot tell whether a person or a robot completed it?

Fan highlighted that progress in embodied AI and physical AI depends on generating large amounts of diverse data, access to open robot foundation models and simulation frameworks — and walked through a unified workflow for developing embodied AI.

With synthetic data workflows like NVIDIA Isaac GR00T-Dreams — built on NVIDIA Cosmos world foundation models— developers can generate virtual worlds from images and prompts, speeding the creation of large sets of diverse and physically accurate data.

That data can then be used to post-train NVIDIA Isaac GR00T N open foundation models for generalized humanoid robot reasoning and skills. But before the models are deployed in the real world, these new robot skills need to be tested in simulation.

Open simulation and learning frameworks such as NVIDIA Isaac Sim and Isaac Lab allow robots to “practice” countless times across millions of virtual environments before operating in the real world, dramatically accelerating learning and deployment cycles.

Plus, with Newton, an open-source, differentiable physics engine built on NVIDIA Warp and OpenUSD, developers can bring high-fidelity simulation to complex robotic dynamics such as motion, balance and contact — reducing the simulation-to-real gap.

This accelerates the creation of physically capable AI systems that learn faster, perform more safely and operate effectively in real-world environments.

However, scaling embodied intelligence isn’t just about compute — it’s about access. Fan reaffirmed NVIDIA’s commitment to open source, emphasizing how the company’s frameworks and foundation models are shared to empower developers and researchers globally.

Developers can get started with NVIDIA’s open embodied and physical AI models on Hugging Face.

NVIDIA’s Llama‑Embed‑Nemotron‑8B model has been recognized as the top open and portable model on the Multilingual Text Embedding Benchmark leaderboard.

Built on the meta‑llama/Llama‑3.1‑8B architecture, Llama‑Embed‑Nemotron‑8B is a research text embedding model that converts text into 4,096‑dimensional vector representations. Designed for flexibility, it supports a wide range of use cases, including retrieval, reranking, semantic similarity and classification across more than 1,000 languages.

Trained on a diverse collection of 16 million query–document pairs — half from public sources and half synthetically generated — the model benefits from refined data generation techniques, hard‑negative mining and model‑merging approaches that contribute to its broad generalization capabilities.

This result builds on NVIDIA’s ongoing research in open, high‑performing AI models. Following earlier leaderboard recognition for the Llama NeMo Retriever ColEmbed model, the success of Llama‑Embed‑Nemotron‑8B highlights the value of openness, transparency and collaboration in advancing AI for the developer community.

Check out Llama-Embed-Nemotron-8B on Hugging Face, and learn more about the model, including architectural highlights, training methodology and performance evaluation.

Open models are shaping the future of AI, enabling developers, enterprises and governments to innovate with transparency, customization and trust. In the latest episode of the NVIDIA AI Podcast, NVIDIA’s Bryan Catanzaro and Jonathan Cohen discuss how open models, datasets and research are laying the foundation for shared progress across the AI ecosystem.

The NVIDIA Nemotron family of open models represents a full-stack approach to AI development, connecting model design to the underlying hardware and software that power it. By releasing Nemotron models, data and training methodologies openly, NVIDIA aims to help others refine, adapt and build upon its work, resulting in a faster exchange of ideas and more efficient systems.

“When we as a community come together — contributing ideas, data and models — we all move faster,” said Catanzaro in the episode. “Open technologies make that possible.”

There’s more happening this week at Open Source AI Week, including the start of the PyTorch Conference — bringing together developers, researchers and innovators pushing the boundaries of open AI.

Attendees can tune in to the special keynote address by Jim Fan, director of robotics and distinguished research scientist at NVIDIA, to hear the latest advancements in robotics — from simulation and synthetic data to accelerated computing. The keynote, titled “The Physical Turing Test: Solving General Purpose Robotics,” will take place on Wednesday, Oct. 22, from 9:50-10:05 a.m. PT.

Computer scientist Andrej Karpathy recently introduced Nanochat, calling it “the best ChatGPT that $100 can buy.” Nanochat is an open-source, full-stack large language model (LLM) implementation built for transparency and experimentation. In about 8,000 lines of minimal, dependency-light code, Nanochat runs the entire LLM pipeline — from tokenization and pretraining to fine-tuning, inference and chat — all through a simple web user interface.

With NVIDIA Launchables, developers can train and interact with their own conversational model in hours with a single click. The Launchables dynamically support different-sized GPUs — including NVIDIA H100 and L40S GPUs — on various clouds without need for modification. They also automatically work on any eight-GPU instance on NVIDIA Brev, so developers can get compute access immediately.

The first 10 users to deploy these Launchables will also receive free compute access to NVIDIA H100 or L40S GPUs.

Start training with Nanochat by deploying a Launchable:

Today, Karpathy received an NVIDIA DGX Spark — the world’s smallest AI supercomputer, designed to bring the power of Blackwell right to a developer’s desktop. With up to a petaflop of AI processing power and 128GB of unified memory in a compact form factor, DGX Spark empowers innovators like Karpathy to experiment, fine-tune and run massive models locally.

PyTorch, the fastest-growing AI framework, derives its performance from the NVIDIA CUDA platform and uses the Python programming language to unlock developer productivity. This year, NVIDIA added Python as a first-class language to the CUDA platform, giving the PyTorch developer community greater access to CUDA.

CUDA Python includes key components that make GPU acceleration in Python easier than ever, with built-in support for kernel fusion, extension module integration and simplified packaging for fast deployment.

Following PyTorch’s open collaboration model, CUDA Python is available on GitHub and PyPI.

Every month, developers worldwide download hundreds of millions of NVIDIA libraries — including CUDA, cuDNN, cuBLAS and CUTLASS — mostly within Python and PyTorch environments. CUDA Python provides nvmath-python, a new library that acts as the bridge between Python code and these highly optimized GPU libraries.

Plus, kernel enhancements and support for next-generation frameworks make NVIDIA accelerated computing more efficient, adaptable and widely accessible.

NVIDIA maintains a long-standing collaboration with the PyTorch community through open-source contributions and technical leadership, as well as by sponsoring and participating in community events and activations.

At PyTorch Conference 2025 in San Francisco, NVIDIA will host a keynote address, five technical sessions and nine poster presentations.

NVIDIA’s on the ground at Open Source AI Week. Stay tuned for a celebration highlighting the spirit of innovation, collaboration and community that drives open-source AI forward. Follow NVIDIA AI Developer on social channels for additional news and insights.

Open Source AI Week kicks off on Monday with a series of hackathons, workshops and meetups spotlighting the latest advances in AI, machine learning and open-source innovation.

The event brings together leading organizations, researchers and open-source communities to share knowledge, collaborate on tools and explore how openness accelerates AI development.

NVIDIA continues to expand access to advanced AI innovation by providing open-source tools, models and datasets designed to empower developers. With more than 1,000 open-source tools on NVIDIA GitHub repositories and over 500 models and 100 datasets on the NVIDIA Hugging Face collections, NVIDIA is accelerating the pace of open, collaborative AI development.

Over the past year, NVIDIA has become the top contributor in Hugging Face repositories, reflecting a deep commitment to sharing models, frameworks and research that empower the community.

Openly available models, tools and datasets are essential to driving innovation and progress. By empowering anyone to use, modify and share technology, it fosters transparency and accelerates discovery, fueling breakthroughs that benefit both industry and communities alike. That’s why NVIDIA is committed to supporting the open source ecosystem.

We’re on the ground all week — stay tuned for a celebration highlighting the spirit of innovation, collaboration and community that drives open-source AI forward, with the PyTorch Conference serving as the flagship event.

The nights grow longer and the shadows get bolder with Vampire The Masquerade: Bloodlines 2 on GeForce NOW, launching with GeForce RTX 5080-power.

Members can sink their teeth into the action role-playing game from Paradox Interactive as part of nine games coming to the cloud this week, including NINJA GAIDEN 4.

Be among the first to play The Outer Worlds 2 with early access available in the cloud starting tomorrow, Oct. 24.

Atlanta is the latest region to get GeForce RTX 5080-class power, with Sofia, Bulgaria, coming next. Stay tuned to GFN Thursday for updates as more regions upgrade to NVIDIA Blackwell RTX. Follow along with the latest progress on the server rollout page.

Vampire: The Masquerade — Bloodlines 2, Paradox’s sequel to the cult classic, invites fans to sink their fangs into Seattle’s dark nightlife, putting players in the immortal boots of Phyre — the newly awakened Elder vampire with a mysterious voice in their head. In every dark corner hides a new alliance or a rival, and every choice carves a path through the bloody politics of the night.

Master Seattle’s secrets and the supernatural abilities of any chosen clan — shadow hunting with the Banu Haqim, blood bending with the Tremere or going fists first as a Brujah. Every story twist and council drama is shaped by dialogue, alliances and whispers from both friends and the ever-present stranger in the protagonist’s mind.

Every neon-lit block streams beautifully on GeForce NOW, with GeForce RTX 5080-class power rolling out for the highest frame rates and sharpest graphics in the cloud — no waiting for downloads or mortal PC specs required.

NINJA GAIDEN 4, a lightning-fast action-adventure title from Team NINJA, slices its way back into the spotlight, packed with brutal, razor-sharp combat and featuring a new protagonist, Yakumo. Known for its intensity, precision and stylish flair, the game demands focus and rewards mastery.

This new entry throws Yakumo into a deadly conflict blending myth and modern chaos. Explore sprawling levels filled with relentless enemies, cinematic boss battles that punish hesitation and a fluid combat system that chains combos together like a storm of steel. Every encounter feels like a duel where speed, timing and finesse define survival.

On GeForce NOW, NINJA GAIDEN 4 reaches its sharpest edge. No installs or massive downloads stand in the way — gain instant access with cloud-powered performance that keeps every slash crisp and every dodge responsive. Whether running on a high-end rig, a laptop or even a mobile device, the action always hits at peak intensity with GeForce NOW.

Early access for The Outer Worlds 2 is reaching the cloud, bringing all the offbeat charm, sharp wit and spacefaring chaos the series is loved for. From Obsidian Entertainment, the masters of branching stories and immersive worlds, this sequel leans even harder into meaningful choices and unexpected consequences.

On GeForce NOW, hopping into early access means instant boarding with no installs or waiting around. Every quip, shootout and twist in the storyline streams smoothly, no matter the screen. It’s the perfect way to jump into space mischief before the game fully launches in the cloud on Wednesday, Oct. 29.

Life finds a way … again. Frontier Developments returns with Jurassic World Evolution 3, the next installment of the park-management series that lets players design, build and wrangle their own prehistoric paradise. Bigger storms, smarter dinosaurs and even bolder decisions keep every moment thrilling — because in this world, control is just an illusion.

In addition, members can look for the following:

What anime would you want to turn into a video game, or vice versa?

— NVIDIA GeForce NOW (@NVIDIAGFN) October 22, 2025

Editor’s note: This blog is a part of Into the Omniverse, a series focused on how developers, 3D practitioners and enterprises can transform their workflows using the latest advances in OpenUSD and NVIDIA Omniverse.

Building robots that can effectively operate alongside human workers in factories, hospitals and public spaces presents an enormous technical challenge. These robots require humanlike dexterity, perception, cognition and whole-body coordination to navigate unpredictable real-world environments in real time.

A “sim-first” approach unlocks these critical skills by enabling parallel training of hundreds or thousands of robot instances using real robot-captured data and synthetic data in simulation environments. Universal Scene Description (OpenUSD) provides the foundational framework for this advanced robot development, serving as a scalable, interoperable data standard that enables developers to build physically accurate virtual worlds where robots can practice and perfect their skills before transferring them to real-world applications.

NVIDIA announced at the Conference on Robot Learning this week groundbreaking advances in open-source physics simulation, open foundation models and development frameworks, including:

Newton Physics Engine: While robots learn faster and safer in simulation, humanoid robots — with complex joints, balance and movements — are pushing today’s physics engines to the limit.

Codeveloped by Google DeepMind, Disney Research and NVIDIA, and managed by the Linux Foundation, Newton is an open-source, GPU-accelerated physics engine to advance robot learning.

Built on NVIDIA Warp and OpenUSD, Newton enables robots to learn complex tasks more precisely while working seamlessly with robot learning frameworks like MuJoCo Playground and NVIDIA Isaac Lab.

Tune in to an upcoming livestream for a Newton beta demonstration to learn about Newton’s core features and how to get started on Newton with NVIDIA Isaac Lab.  

Isaac GR00T N1.6: To perform humanlike tasks in the physical world, humanoids must understand ambiguous instructions and navigate unforeseen scenarios. The latest Isaac GR00T N1.6 open robot foundation model, available soon on Hugging Face, integrates NVIDIA Cosmos Reason, an open reasoning vision language model built for physical AI. Cosmos Reason serves as the robot’s deep-thinking brain and transforms vague instructions into step-by-step action plans using prior knowledge, common sense and physics understanding.

NVIDIA Isaac Lab: The latest version of Isaac Lab, an open-source, modular robot learning framework built on NVIDIA Isaac Sim and OpenUSD, is now available as an early developer release. Version 2.3 brings a host of new features to robotics researchers and developers, including advanced whole-body control and expanded teleoperation for data collection.

OpenUSD’s interoperability ensures these advanced physics simulations, foundation models and learning frameworks work together seamlessly, enabling developers to build unified robot learning pipelines that scale across different platforms and deployment scenarios.

Leading humanoid and robotics developers, including Agility Robotics, Lightwheel, Mentee, Universal Robots and Wandelbots are adopting simulation technologies and libraries to accelerate physical AI development and deployment.

NVIDIA’s open frameworks and libraries are also being adopted by developers within the robotics community. Community member and NVIDIA Omniverse ambassador Dylan Tobin created an AI chatbot trained on Isaac Sim workflows to help developers navigate Omniverse more efficiently.

See how other developers in the community are using Isaac Sim and Isaac Lab for innovation in robotics navigation, control and reinforcement learning by watching this livestream replay:

Plus, an NVIDIA Robotics office hours session demonstrates how Brev makes it easy to run Isaac Sim and Isaac Lab on Omniverse:

To learn more about robot learning with OpenUSD and NVIDIA’s latest robotics technologies, explore these resources:

Stay up to date by subscribing to NVIDIA Omniverse news, joining the Omniverse community and following Omniverse on Discord, Instagram, LinkedIn, Threads, X and YouTube. 

Explore the Alliance for OpenUSD forum and the AOUSD website.