TechPrufer
Get started
NVIDIA
NVIDIA logoNVIDIA · Feature trail

Agentic AI Infrastructure Acceleration with BlueField DPUs

72 posts · May 2026 – Oct 2026

What this trail covers

This feature thread tracks the evolution of infrastructure designed to support agentic AI workloads, particularly focusing on the challenges and advancements in large-scale model pre-training and deployment. Initial efforts focused on agentic AI introducing complex distributed workflows requiring tight coupling of GPUs, CPUs, memory, networking, storage, and security. This demanded infrastructure capable of moving, protecting, retrieving, and reusing context data at high speeds without latency. Subsequent developments have focused on building reliable data analytics agents, emphasizing constrained harnesses, unified data access, schema context seeding, opinionated toolsets, and effective handling of structured and unstructured data sources like documents and videos. Techniques include preprocessing data, limiting the agent's action space, and making every attempt inspectable through execution traces and specialized agents for failure analysis. This extends to optimizing the interaction between agents and underlying infrastructure by focusing on the harness as a key optimization surface, especially when working with smaller, fixed LLMs.

All NVIDIA trails
  1. Building Reliable Data Analytics Agents: Lessons from the KDD Cup | NVIDIA Technical BlogThis post details the KGMON team's approach to building reliable data analytics agents for the KDD Cup 2026 competition. Key contributions include: 1. Unifying structured data sources (databases, CSV, JSON) into a single SQLite database with a narrow schema() and sql() interface to reduce routing errors and preserve reasoning capacity. 2. Implementing a preflight schema-scouting step to brief the agent on tables, join keys, and data-quality issues before the main reasoning loop. 3. Building a small, opinionated tool harness with functions like schema(), sql(query), write_answer(df), and prose_helper(), with middleware for repairing malformed calls and persistent state for intermediate results. 4. Separating document inspection from structured analysis by using targeted previews and regex searches fed to a prose_helper tool (with zero-temperature LLM calls) to extract answers or tables, keeping raw prose out of the main context. 5. Preprocessing video data by extracting keyframes, transcribing audio, and aligning segments to provide evidence to the agent. 6. Logging execution traces (prompts, tool calls, intermediate results) for a specialized inspector agent to categorize failures and guide harness improvements. The core lesson is that reliability often comes from the harness design around the model, especially for smaller LLMs.
  2. Validate AI Factory Changes with Digital Twins and AI Agents | NVIDIA Technical BlogThis post introduces NVIDIA DSX Air as a node-based digital twin simulation for AI factory infrastructure and operational interfaces, enabling pre-production validation of hardware topology and software stack changes. It highlights the integration of agentic AI workflows to interrogate the twin, run configuration checks, compare outcomes against policy, and produce evidence-backed recommendations. NVIDIA Brev is presented as a complementary service providing on-demand GPU compute that connects to DSX Air for executing validated tasks against the simulated factory. The NVIDIA AI Blueprint for Video Search and Summarization is used as a demonstration of this pattern, combining video analysis, retrieval-augmented knowledge, and agent orchestration within the digital twin.
  3. Scaling Decision Optimization to 100 Million Variables and Beyond with mPDLP in NVIDIA cuOpt | NVIDIA Technical BlogThis post introduces the NVIDIA cuOpt Multi-GPU Primal-Dual hybrid gradient for Linear Programming (mPDLP) solver. It details how mPDLP distributes LP problems across NVLink-connected GPUs, reducing solve time for large problems and achieving up to 6x lower peak memory usage per GPU compared to single-GPU PDLP. The post explains the PDLP algorithm, the challenges of distributing it (communication overhead and load imbalance), and introduces min-cut partitioning as an improvement over 2D partitioning by exploiting shared dependencies between consecutive Sparse Matrix-Vector Multiplications (SpMVs). Benchmarking results and partner applications (Kinaxis, PSR) are presented to demonstrate the performance gains.
  4. Build Applications on NVIDIA BlueField Faster with NVIDIA DOCA Agent Skills | NVIDIA Technical BlogThis post introduces NVIDIA DOCA AI agent skills, a new capability that provides structured, verified foundations for AI agents developing on NVIDIA BlueField DPUs. These skills include verified API signatures, hardware capability requirements, and build constraints, addressing the limitations of general-purpose agents. The post quantifies the benefits through a 65-prompt evaluation, showing a significant improvement in agent performance (from 19% to 100% checklist item satisfaction). It details four key benefits: faster development with real APIs, hardware verification before coding, compilation and linking correctness, and safe execution of hardware-affecting changes. A demo illustrates a 73% reduction in handwritten code and 46% fewer hardware commands for an RDMA application.
  5. Tracing Agent Harness Behavior with NVIDIA NeMo Relay | NVIDIA Technical BlogThis post introduces NVIDIA NeMo Relay as a common observability layer for agent harnesses, enabling detailed tracing of agent behavior. It demonstrates how NeMo Relay integrates with Hermes Agent to produce ATOF event streams, ATIF trajectories, and OpenTelemetry spans. The post provides a practical example of using these traces to debug agent inefficiencies, compare harness revisions (e.g., Hermes ToolPerf case study), and verify task outcomes. It highlights the value of structured traces for understanding model and tool interactions, token usage, and error recovery, contributing to the development of more robust and auditable AI agents.
  7. Lower the Cost of Building and Running Visual AI Agents with NVIDIA VSS Blueprint 3.3 | NVIDIA Technical BlogVSS Blueprint 3.3 introduces the Build Vision Agent skill (`vss-build-vision-ai`) which composes multi-workflow deployments from a single prompt, starting from one of four validated developer profiles and adding only the services a requested capability actually reaches while converging shared infrastructure such as Kafka, Redis, and Elasticsearch onto single instances. It automates mapping application goals to VSS workflows, combining workflows, reusing shared infrastructure, generating self-contained deployments, showing architecture diagrams, and extending running deployments. Adaptive Efficient Video Sampling (EVS) reduces runtime VLM processing by dynamically pruning unchanged visual patches per frame and batching VLM work around moments of activity, cutting alert contextualization latency by 17% and increasing concurrent real-time VLM streams by 46% on an RTX PRO 6000 Blackwell running Cosmos 3 Super FP8.
  8. How SWE-Serve Exposes the Gap Between Local Tests and Live Serving | NVIDIA Technical BlogThis post introduces SWE-Serve, a new benchmark for evaluating AI coding agents on inference-engineering tasks. It highlights the critical gap between local testing and live serving of AI models by demonstrating that a substantial percentage of patches passing other tests fail when validated through a live serving interface. The benchmark comprises 53 tasks derived from 83 merged SGLang pull requests, covering model enablement, decoding, caching, scheduling, serving APIs, and distributed execution. SWE-Serve reveals that tasks spanning multiple runtime domains (request handling, scheduling, model execution, KV-cache management) have a significantly lower pass rate compared to single-domain tasks. The post also presents performance data for 11 AI models across these tasks, showing wide variations in pass@1 scores, cost, and wall time, and notes that no single model excels across all task categories.
  9. Accelerating a ROS 2 Node with an AI Agent and NVIDIA Isaac ROS | NVIDIA Technical BlogThis post details the integration of `rosidl::Buffer` with a CUDA buffer backend into a ROS 2 node (Depth Anything 3 TensorRT) to enable zero-copy transport of GPU-resident payloads. It introduces an AI coding agent skill (`migrate-node-to-rosidl-buffer`) to automate the audit, planning, and refactoring process for adopting this capability. The contribution lies in demonstrating an agentic workflow for optimizing inter-node data movement in robotics AI, preserving existing ROS contracts while leveraging GPU acceleration more effectively.
  10. How to Evaluate AI Agents From Tool Calls to Task Completion | NVIDIA Technical BlogThis post details a new framework for evaluating AI agents, shifting focus from individual tool call accuracy to full task completion in stateful environments. It introduces a two-layer scoring system (step-level process scoring and end-to-end outcome scoring) and a hierarchical metric structure (Benchmark → Trial → Task → Turn → Step) with axes for accuracy, verbosity, and cost. Key metrics include task success rate, consistency, tool-call precision, argument accuracy, steps per success, and cost per success. The post emphasizes the importance of benchmark comparability through task complexity, environment statefulness, and executable verification methodologies, while also highlighting the risks of data contamination. The Nemotron 3.5 Lightning model is presented as an example demonstrating this evaluation framework.
  11. From Wafer-Out to First Token: Codifying Supply Chain Expertise with Nemotron and Palantir Foundry | NVIDIA Technical BlogThis post details the integration of Nemotron 3.5 Lightning, post-trained on supply chain allocation decisions, with Palantir Foundry and NVIDIA cuOpt. It describes how captured allocation decisions, rationales, and outcomes were used to train a specialized 30B model that achieved 86.7% accuracy on a development benchmark. The workflow captures human expertise to improve quantitative optimization and accelerate the time-to-token for complex hardware supply chains.
  12. Building a Memory-Driven Agent with NVIDIA NemoClaw | NVIDIA Technical BlogThis post introduces NVIDIA NemoClaw for building memory-driven AI agents. It details the 'self model' concept for maintaining structured knowledge (people, projects, priorities) in Markdown and a SQLite ledger for obligations and judgments. Key contributions include separating evidence from judgment, prioritizing user intent over urgency via an 'intent gate', enabling user corrections through an append-only audit trail, and enforcing runtime security with NVIDIA OpenShell. Evaluation metrics demonstrate improved accuracy and fact tracking compared to a RAG baseline.
  13. Building an Adaptive Agentic Cybersecurity System with NVIDIA Nemotron | NVIDIA Technical BlogThis post details the evaluation of an adaptive agentic cybersecurity system combining NVIDIA Nemotron open models with CrowdStrike's Falcon telemetry and agent harnesses. It describes the closed-loop offense-defense testing methodology, the creation of a representative test environment based on NVIDIA's accelerated computing infrastructure, and the specialization of the defensive harness with mechanisms for schema knowledge, telemetry grounding, specialized detection authoring (using Nemotron 3 Super), artifact linting, detection replay, and independent review. The evaluation demonstrated significant improvements in detection rates and generalization compared to baseline systems, highlighting the potential of agentic AI for cybersecurity operations.
  14. Run NVIDIA BioNeMo NIM Microservices for Protein Structure Prediction in Claude Science | NVIDIA Technical BlogThis post details the integration of the NVIDIA BioNeMo Agent Toolkit with Claude Science, enabling AI agents to orchestrate protein structure prediction workflows. It demonstrates the use of NIM microservices for MSA generation (MSA Search NIM) and protein folding (OpenFold3 NIM, Boltz-2 NIM). The post outlines the setup process, including connecting to compute endpoints and launching local NIM containers, and provides specific prompts for generating MSAs and performing structure predictions for both monomer and heteromer cases. It highlights the importance of MSA input for accurate interface prediction and presents a comparative analysis of OpenFold3 and Boltz-2 predictions.
  15. How to Train a Cross-Embodiment Robot Navigation Policy with AI Agents | NVIDIA Technical BlogThis post introduces the COMPASS framework and an agent-driven workflow for training cross-embodiment robot navigation policies. It details the use of residual reinforcement learning to adapt a pre-trained X-Mobility policy to specific robots and environments, leveraging tools like NVIDIA Omniverse NuRec for scene reconstruction and simulation. The workflow includes automated environment validation, scene preparation, smoke testing, residual training, checkpoint evaluation, and runtime packaging, with human approval gates.
  16. Experiment with Qwen3.8-Flash-Next on NVIDIA GB300 NVL72 for Agentic Coding | NVIDIA Technical BlogThis post details the experimental use of Qwen3.8-Flash-Next, a multimodal MoE model with a 125B parameter main model and a 262,144-token context window (extensible to 1M tokens), on the NVIDIA GB300 NVL72 platform for agentic coding applications. It highlights the model's architectural innovations, Gated DeltaNet (GDN) and Qwen Sparse Attention (QSA), which address KV cache growth and attention bottlenecks for long contexts, achieving significant speedups over full attention. The post quantifies performance on the GB300 NVL72, reporting over 16K tokens/sec/GPU and over 200 tokens/sec/user. It also mentions fine-tuning with NVIDIA NeMo AutoModel and reinforcement learning with NeMo RL recipes, and provides integration details with inference engines like SGLang, vLLM, and TokenSpeed.
  17. NVIDIA Vera Rubin and Blackwell Set a New Standard for Agentic AI Performance per Watt | NVIDIA Technical BlogThis post introduces NVIDIA Vera Rubin NVL72 and its performance improvements for agentic AI workloads, specifically using the SemiAnalysis AgentX benchmark. It details how Vera Rubin NVL72 achieves up to 30x higher AI-factory throughput per megawatt than GB300 NVL72 at 160 tokens per second per user on the AgentX DeepSeek V4-Pro workload. The post also highlights system-level optimizations such as MoE serving runtimes (SGLang, TensorRT-LLM, vLLM), DeepGEMM-based kernels with MXFP4 and MXFP8 formats, NVIDIA Dynamo for disaggregated prefill and decode, and NVLink scale-up fabric, which collectively enable responsive multi-turn serving for agent sessions.
  18. How NVIDIA Groq 3 LPX Unlocks Ultrafast Interactivity at Long Context on NVIDIA Vera Rubin | NVIDIA Technical BlogThis post introduces the NVIDIA Groq 3 LPX as an interactive AI inference accelerator for the NVIDIA Vera Rubin platform, enabling ultrafast interactivity at long context lengths (100K+ tokens) for agentic AI. It details how Groq 3 LPX achieves this through a deterministic execution model with compiler-scheduled chip-to-chip communication and computation-communication overlap, minimizing first-bit latency. Benchmarks show median speeds of 3,431 output tokens/second on Gemma 4 31B at 100K context and 4,767 tokens/second on the SPEED-Bench coding benchmark. The post also discusses configurations like prefill-decode disaggregation and attention-FFN disaggregation for optimizing performance.
  19. NVIDIA BlueField-4 Powers New Scale-In Network Infrastructure for Agentic AI Factories | NVIDIA Technical BlogThis post introduces 'Scale-In' network infrastructure as the fifth pillar of NVIDIA AI networking, specifically detailing the role of NVIDIA BlueField-4, DOCA, and Spectrum-X Ethernet in accelerating the north-south access path for agentic AI factories. It explains how BlueField-4's host-independent processing, 800 Gb/s interface, and inline acceleration engines enable line-rate networking, storage, and security. The post highlights DOCA's role in programming these services via containerized microservices and SDKs for unified operation, and Spectrum-X Ethernet's contribution to high-performance fabric with congestion management. It also outlines specific use cases like isolated AI factory VPCs, hardware-enforced security, accelerated storage, DPU-native orchestration, and fleet-wide observability. The integration of BlueField Astra for unified control across Scale-In and Scale-Out domains is also discussed.
  20. Solving Agentic AI Fleet Challenges with NVIDIA Vera CPU | NVIDIA Technical BlogThis post introduces the NVIDIA Vera CPU, architected with NVIDIA Olympus cores, to address agentic AI fleet challenges. Telemetry from 163,594 agentic sessions reveals that over 97% have unique trajectory profiles, making specialized CPU designs impractical. Agentic workloads exhibit a long sequential chain of reasoning with sporadic parallel bursts, where the sequential path is latency-bound and dictates overall session completion time. The Vera CPU balances strong per-thread performance for the critical path with concurrency for fan-out bursts. Estimated SPEC CPU 2026 results show up to 1.5x per-core performance improvement over AMD Venice across agentic workloads like compiler, static analysis, and Python benchmarks. The balanced design maximizes completed user sessions by efficiently utilizing cores, memory bandwidth, and power, avoiding stranded resources and the need for a fragmented fleet.
  21. NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents | NVIDIA Technical BlogThis post introduces the NVIDIA Agentic Variation Operators (AVO) architecture, a general-purpose agent system designed for sustained autonomous operation across long horizons. It details AVO's application to GPU-kernel optimization, where it achieved significant performance gains over existing methods by autonomously exploring and refining kernels. The post then describes the adaptation of the same AVO architecture to the ARC-AGI-3 interactive reasoning benchmark, where it achieved a 100.00 RHAE score. Key contributions include demonstrating that long-horizon capability emerges from the full agent system (memory, tools, feedback, recovery) rather than solely from model capability, and highlighting the transferability of the core agent loop across different domains (software engineering and interactive reasoning). The post emphasizes the importance of persistent memory and supervision for maintaining progress and adapting strategies.
  22. Where Security Fits in an AI Agent Stack | NVIDIA Technical BlogThis post details the security considerations for AI agent stacks, focusing on the distinction between behavioral and infrastructure controls. It introduces NVIDIA OpenShell as a secure runtime layer that enforces isolation, identity, policy, credentials, and audit for autonomous agents, and NVIDIA Dynamo as the inference data plane. Five design rules are proposed for enforceable agent security: higher layers propose actions while lower layers decide, policy remains below the security boundary, every effect is checked, access is just-in-time, and isolation enables recovery. Four security profiles (Isolated, Connected, Production, and Adversarial) are outlined, applying progressively stricter controls. The post emphasizes establishing the AI agent runtime boundary before launch and highlights common security gaps such as unclear boundaries, excessive access, untrusted data as control, uncontrolled external effects, compounding failures, and incomplete audit evidence.
  23. Developing NVIDIA Holoscan Applications with CLI, Skills, and AI Coding Agents | NVIDIA Technical BlogThis post details the iterative development of a real-time endoscopic tool segmentation application using NVIDIA Holoscan, HoloHub, and an AI coding agent. It demonstrates how an AI agent, guided by an engineer, can leverage the Holoscan CLI, HoloHub documentation, and development skills to create, test, and benchmark applications. Key contributions include the establishment of an engineer-guided, agentic development workflow, the creation of a minimal working application, the implementation of repeatable benchmarking modes (visual, smoke, benchmark) using Holoscan Data Flow Tracking, and an ablation study showing the effectiveness of combining CLI, skills, and documentation for agentic development.
  24. Evaluating AI Agent Skill Performance with NVIDIA SkillEvaluator | NVIDIA Technical BlogThis post introduces NVIDIA SkillEvaluator, an open-source tool for measuring the impact of AI agent skills on agent performance. It details a three-tier evaluation methodology: static checks (schema validation, security scanning, PII detection), distinctiveness analysis (embedding similarity for overlap), and live task runs in isolated environments using the Harbor framework. The post shares the first benchmark results for over 300 verified skills across 30+ NVIDIA products, showing significant Skill Lift (average +31 points, +39 excluding security) in Correctness, Discoverability, Effectiveness, and Efficiency when skills are used. It also notes that product domain and evaluation design are more influential on Skill Lift than the agent harness choice.
  25. Serve Qwen3.8-2.4T-A95B, a 2.4T-Parameter Model, with Configurable Reasoning on NVIDIA GB300 NVL72 | NVIDIA Technical BlogThis post details the deployment and performance optimization of the Qwen3.8-2.4T-A95B, a 2.4T-parameter Mixture of Experts (MoE) model, on the NVIDIA GB300 NVL72 platform. It highlights architectural innovations for long-context inference, including a hybrid full/linear attention mechanism and fine-grained MoE, which enable efficient handling of up to one million tokens. The post presents performance metrics (over 4K tokens/sec/GPU, 350 tokens/sec/user) achieved with FP8 precision using TensorRT-LLM and discusses various open-source inference recipes (SGLang, vLLM, NVIDIA Dynamo) and deployment options (NVIDIA NIM). It also mentions post-training capabilities with NVIDIA NeMo AutoModel and LoRA fine-tuning.
  26. NVIDIA JetPack 7.2.1 Adds Agentic Video Skills and T3000 Emulation | NVIDIA Technical BlogThis post introduces foundational agentic video skills within NVIDIA JetPack 7.2.1, building upon existing SDKs like Video Codec SDK and PyNvVideoCodec. These skills enable programmable, device-aware video workflows by connecting developer intent to live device discovery, supported configurations, reproducible execution, and performance verification. The skills allow for independent invocation or combination to configure and verify codec stages within broader pipelines, including identifying platform capabilities, generating encoder recipes, benchmarking performance, and validating the codec workflow. This represents an agentic layer above the video primitives provided by the SDKs, focusing on device-aware configuration and verification for video processing in AI applications.
  27. NVIDIA Nemotron 3.5 Lightning Delivers Fast, Accurate Specialized Task Execution for Long-Running Agents | NVIDIA Technical BlogThis post introduces NVIDIA Nemotron 3.5 Lightning, a 30B parameter open Mixture-of-Experts (MoE) model with 3B active parameters, optimized for high-volume, low-latency execution in always-on AI agents and agentic workflows. It details innovations in inference and training, including speculative decoding, harness-optimized training, and quantization (NVFP4 and BF16 checkpoints), delivering up to 4x output speed compared to similar-sized models. The post also introduces NVIDIA NeMo Switchyard, a library for intelligent model routing, allowing Nemotron 3.5 Lightning to be deployed alongside other models for optimal task allocation. The open release includes permissive licensing, weights, data, and recipes for customization.
  28. Route AI Agents Across Models with NVIDIA NeMo Switchyard | NVIDIA Technical BlogIntroduces NVIDIA NeMo Switchyard, an orchestration layer for AI agents that enables dynamic routing of workloads across specialized and frontier models. It provides a provider-agnostic SDK, supports tuning-free and tunable routing algorithms, and separates routing logic from model providers. The post details how Switchyard uses signals from model capabilities, cost profiles, and infrastructure to make routing decisions, and includes architectural diagrams and benchmark results showing cost savings.
  29. Run Local Agentic AI Workflows with Meta’s Muse Glimmer on NVIDIA | NVIDIA Technical BlogThis post introduces Meta's Muse Glimmer, a 30B parameter dense model with a 120K+ context window, optimized for local, long-running agentic AI workflows on NVIDIA GPUs. It details the model's dense architecture for reliability and predictable latency, its ability to run fully on-device across various NVIDIA platforms (GeForce RTX 5090, DGX Spark, DGX Station, Jetson), and its performance metrics (20K tokens/sec/GPU on Blackwell Ultra). It also highlights integration with NVIDIA NemoClaw for agent scaffolding, NeMo AutoModel for fine-tuning, and flexible deployment paths via NVIDIA NIM, SGLang, and vLLM.
  30. Generate Trajectories, Reasoning Traces, and Auto-Labels with NVIDIA Alpamayo 2 Super | NVIDIA Technical BlogIntroduced NVIDIA Alpamayo 2 Super, a 34B parameter reasoning vision-language-action model that unifies trajectory generation, reasoning trace generation, meta-action prediction, VQA with 2D grounding, and auto-labeling for AV workflows. The model processes 360-degree perception from up to seven cameras and outputs future trajectories, Chain-of-Causation reasoning traces, high-level meta-actions, grounded scene answers, and structured reasoning auto-labels. It achieves state-of-the-art results in trajectory prediction, AV reasoning, LingoQA, meta-action IoU, VQA answer similarity, 2D grounding IoU, and closed-loop AlpaSim Score.
  31. Four Ways to Deploy More Secure AI Agents | NVIDIA Technical BlogThis post details four key security controls for deploying AI agents at enterprise scale: implementing robust access controls, limiting code execution capabilities, enforcing a default-deny network egress policy with least-privilege allowlists, and keeping secrets out of the agent's reach. It highlights the limitations of prompt-based and LLM-as-a-judge defenses against adversarial techniques and emphasizes the need for deterministic architectural controls.
  32. NVIDIA Exemplar Cloud: Lessons for Unlocking Full Performance on AI Infrastructure | NVIDIA Technical BlogThis post details four case studies of performance degradation in AI training infrastructure, identifying specific configuration gaps at the kernel, hypervisor, BIOS, and NCCL levels. It provides diagnostic patterns using tools like perf, Nsight Systems, and NCCL tests to pinpoint issues such as SMMU overhead in virtualized environments (GB200 NVL72), CPU power management and NUMA misbinding (H100 cluster), insufficient NCCL queue-pair concurrency on high-bandwidth fabrics, and missing runtime topology files within containers. The post outlines the fixes applied, including enabling CMDQV for SMMU, optimizing CPU C-states and NUMA bindings, tuning NCCL settings, and ensuring correct topology propagation, leading to significant improvements in training throughput.
  33. How to Self-Host a Validated AI Coding Assistant with NVIDIA NeMo Guardrails | NVIDIA Technical BlogThis post introduces the integration of NVIDIA NeMo Guardrails, a CI verification gate, and an outcome metrics loop (Prometheus/Grafana) to create a self-hosted, validated AI coding assistant. It details the deployment of StarCoder2-7B NIM, the integration of NeMo Guardrails for policy enforcement (e.g., blocking requests for human-only files), and the setup of a CI pipeline to scan for hallucinated dependencies. The post also outlines how to establish a metrics loop to track the impact of AI-assisted changes on defect rates, ensuring that policy enforcement, dependency verification, and outcome measurement are external to the model itself.
  34. NVIDIA Ising Enables Fully Automated Quantum Computer Calibration with Enhanced In-Context Learning | NVIDIA Technical BlogIntroduces NVIDIA Ising Calibration 1.5, a 31B-parameter vision language model for diagnosing and tuning quantum processors. This release features a new NVFP4-quantized version for single GPU or NVIDIA DGX Spark deployment, reducing model size by 11.4% at BF16 precision. The model is trained on diverse datasets from multiple qubit modalities and evaluated with the QCalEval benchmark, demonstrating state-of-the-art zero-shot and in-context learning performance. Full-parameter checkpoints, quantized versions, open datasets, and deployment blueprints are available under the OpenMDW License, with optimized tokens per second throughput on NVIDIA DGX Spark and integration support through the NVIDIA Nemo Agent Toolkit for automated quantum calibration workflows.
  35. Six Agent Harness Capabilities for Higher Model Performance | NVIDIA Technical BlogThis post introduces NVIDIA Labs Object-Oriented Agents (NOOA), an open-source research preview agent framework. NOOA structures agents as single Python classes, integrating capabilities, state, and prompts via methods, fields, and docstrings, with type annotations as enforced contracts. LLM-driven loops complete method bodies marked by ellipses at runtime. The NOOA harness enables agents to curate and persist long-term, typed, and relational memory in a human-readable SQLite store, supporting knowledge accumulation, efficient context management through pass-by-reference, and eliminating the need for context compaction or summarization pipelines. Benchmarks on SWE-bench Verified, CyberGym L1, and ARC-AGI-3 demonstrate state-of-the-art performance and efficiency, achieving higher accuracy and lower token cost compared to prior harnesses, with reproducible, inspectable workflows and open evaluation methodology. Key performance drivers identified are typed input/output, pass by reference, code as action, programmable loop engineering, explicit object state, and model-callable harness APIs.
  36. NVIDIA Nemotron 3 Ultra Leads Open Models on Accuracy and Efficiency in Agentic RTL Coding | NVIDIA Technical BlogThis post details the combination of the ACE-RTL agent and NVIDIA Nemotron 3 Ultra for agentic RTL coding. It demonstrates how Nemotron 3 Ultra, with its hybrid Mamba-Attention MoE architecture and 1M-token context length, achieves 97.1% average pass rate on the CVDP benchmark with significantly lower token usage (28% less than GLM 5.2, 71% less than Kimi K2.6). The post highlights the importance of iterative feedback and context evolution for RTL tasks and showcases Nemotron 3 Ultra's efficiency for long-running agentic reasoning due to its architecture and training on synthetic RTL data.
  37. ModelExpress: Distributing Model Artifacts at the Speed of Light | NVIDIA Technical BlogThis post introduces NVIDIA ModelExpress (MX), a system designed to accelerate the model weight lifecycle by optimizing the transfer of model artifacts. MX prioritizes direct GPU-to-GPU P2P RDMA transfers via NIXL, bypassing object storage and host memory. It implements strategies for bootstrapping from storage (Model Streamer, GPUDirect Storage, Model Cache Service) and for scaling out via peer-to-peer transfers once a replica is serving. MX also optimizes NIXL memory registration overhead through techniques like VMM arena registration and efficient memory registration for P2P transfers.
  38. Setting a World Record for MoE Pre-Training on NVIDIA GB300 NVL72 | NVIDIA Technical BlogThis post details a world record set for pre-training the DeepSeek-V3 671B Mixture of Experts (MoE) model on the NVIDIA GB300 NVL72 system. It highlights the critical role of MoE architectures in pushing frontier AI capabilities and the associated communication challenges. The post emphasizes how the GB300 NVL72, through extreme co-design of silicon, interconnect (fifth-generation NVLink), networking (ConnectX-8 SuperNICs, Quantum-X800 InfiniBand/Spectrum-X Ethernet), and software (Megatron Core, TorchTitan, JAX), addresses these challenges. It showcases a 3x performance improvement over previous generations (GB200 NVL72) and a 1.5x gain in six months due to software optimizations, underscoring the continuous evolution of AI infrastructure for large-scale training.
  39. Inside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI | NVIDIA Technical BlogThis post introduces the NVIDIA Rubin GPU architecture and the Vera Rubin platform, detailing hardware innovations that accelerate agentic AI workloads. Key contributions include: 1. Rubin GPU architecture: 336 billion transistors, 224 SMs, 896 Tensor Cores with expanded precision, third-generation Transformer Engine (up to 50 petaflops NVFP4), 288 GB HBM4 memory (22 TB/s bandwidth), NVLink 6, NVLink-C2C, PCIe Gen 6, and Confidential Computing. 2. MoE acceleration: Inline descriptor updates for Tensor Memory Accelerator (TMA) simplify descriptor sharing and reduce data-movement overhead for Mixture-of-Experts models. 3. K-dimension optimization: Doubling the K-dimension processing per clock reduces loop iterations and improves Tensor Core utilization for both context and decode GEMMs at high tensor-parallel scale. 4. Rack-scale integration: The Vera Rubin NVL72 features liquid cooling, DSX MaxLPS for power management, cable-free MGX architecture, and hot-swappable NVLink switch trays for dense, resilient AI supercomputing.
  40. NVIDIA Vera CPU: Olympus Cores Built for Maximum Single-Thread Performance in Agentic AI | NVIDIA Technical BlogThis post introduces the NVIDIA Vera CPU and its Olympus core, designed to accelerate agentic AI workloads by maximizing single-thread performance. It details the Olympus core architecture, including its front-end (optimized branch predictors), mid-core (dependency breaking), execution engine (complex vector optimization), and cache subsystem (latency-optimized paths). It also highlights multithreading innovations (Spatial Multithreading), the Scalable Coherency Fabric, SOCAMM2 memory modules, and secure scale-up connectivity (NVLink-C2C, PCIe 6.4, CXL 3.1, Confidential Computing) to support predictable agent workload throughput and secure VM isolation.
  41. NVIDIA NVLink: The Scale-Up Network for AI Factories | NVIDIA Technical BlogThis post introduces the Sixth Generation NVIDIA NVLink as a critical component for AI factories, highlighting its role as a purpose-built, high-bandwidth, low-latency scale-up networking fabric. It details how NVLink's extreme co-design approach, integrating hardware and software, enables features like disaggregated inference and expert parallelism, leading to significant performance gains in large-scale MoE and LLM workloads. The post also emphasizes NVLink's contribution to factory resiliency, operational features, and its importance in driving sustained ROI by optimizing delivered tokens per watt, dollar, and factory footprint.
  42. Integrate NVIDIA Omniverse RTX Sensor Simulation Into Existing Apps | NVIDIA Technical BlogThis post introduces the integration of NVIDIA Omniverse RTX Sensor Simulation (ovrtx) into existing applications as part of the NVIDIA Agent Toolkit. It highlights how ovrtx provides modular APIs for generating camera, lidar, and radar sensor outputs from OpenUSD scenes, enabling real-time, physically grounded data for applications in design, robotics, and industrial digital twins. The post details the integration process, emphasizing how developers can maintain control over their application architecture while leveraging Omniverse libraries for synthetic data generation, perception testing, and physical AI validation.
  43. Integrating Context-Aware Video AI Agents Into Enterprise Workflows | NVIDIA Technical BlogThis post details the integration of context-aware video AI agents into enterprise workflows using NVIDIA NemoClaw, NVIDIA Blueprints for Video Search and Summarization (VSS), and Retrieval-Augmented Generation (RAG). It explains how to extend VSS for guided, context-aware video analysis, orchestrate VSS and RAG as a composable service with NemoClaw, generate structured reports enriched with organizational and reference knowledge, build multi-step workflows where video analysis feeds into other business processes, and deploy and scale this solution across enterprise environments. The post demonstrates a 'healthy eating coach' example where a user uploads a food video, specifies analysis parameters via human-in-the-loop prompts, NemoClaw orchestrates VSS and RAG to retrieve nutritional guidelines and analyze the video, and then generates a structured report and a Jira ticket with recommended actions.
  44. Scaling Agentic AI Factories Through Extreme Co-Design with NVIDIA BlueField | NVIDIA Technical BlogThis post details how agentic AI and long-context inference drive new infrastructure demands, and how BlueField-4, Vera BlueField-4 STX, and DOCA address them by offloading, accelerating, and isolating infrastructure services across the AI factory data path. It highlights the BlueField-4 DPU's capabilities including 800 Gb/s networking, a 64-core Grace CPU, PCIe Gen6, and LPDDR5X memory, and the Vera BlueField-4 STX Storage Processor's role in AI-native storage and KV cache management. DOCA software is emphasized for providing a programmable foundation for context reuse, zero-trust security, and multi-tenant lifecycle management.
  45. Lessons From the Leaderboard: What 5,000+ Kagglers Taught Us About Improving AI Reasoning | NVIDIA Technical BlogThis post contributes practical lessons learned from the Kaggle Nemotron Model Reasoning Challenge, which involved over 5,000 participants optimizing AI reasoning. It details five key techniques for improving reasoning accuracy: 1. Verifiable Chain-of-Thought Data: Emphasizing the importance of checking intermediate reasoning steps rather than just final answers, using solvers, rule checkers, or human review. 2. Token Budget Design: Strategies for compressing reasoning traces to fit within token limits without losing logical signal, applicable to passing long prompts, retrieval results, or tool outputs. 3. Separating Memory from Solving: Storing reusable knowledge (schemas, formulas, patterns) separately from the live reasoning task to reduce the model's cognitive load. 4. Tool Use and Verification: Demonstrating how top solutions used tools to generate and audit high-quality training data and verified intermediate steps. 5. Community-Driven Workflow Optimization: Highlighting the value of community discussions for sharing techniques, comparing failures, and surfacing edge cases. The post also notes the use of specific infrastructure (Google Cloud G4 VMs with NVIDIA RTX PRO 6000 Blackwell GPUs) and constraints (no internet access at evaluation, LoRA adapters) that shaped the learned techniques.
  46. Reducing High-Bandwidth Memory Bottlenecks in JAX-Based LLM Training with Host Offloading | NVIDIA Technical BlogIntroduces host offloading in JAX for LLM training to reduce HBM bottlenecks by moving activations to host memory. Demonstrates significant throughput improvements (up to 57%) and increased feasible batch sizes on NVIDIA GB200 NVL72 systems using MaxText with Llama 3.1 405B and DeepSeek-V3 671B models. Highlights the importance of co-design between XLA compiler and NVIDIA hardware (NVLink-C2C) for efficient asynchronous data movement and performance gains, especially for sparse MoE models.
  47. Synthetic Data Generation for Financial AI Research with NVIDIA NeMo | NVIDIA Technical BlogThis post details the engineering of an iterative synthetic data generation pipeline for financial AI research. It describes the architecture and implementation of a system that uses NVIDIA NeMo Data Designer for structured generation, NeMo Curator for scalable semantic deduplication (with specific parameters for embedding models, similarity thresholds, and clustering), and NVIDIA Nemotron models served via vLLM (with details on tensor parallelism and concurrent requests). The post elaborates on the iterative process, including category-weighted sampling, filtering malformed outputs, global semantic deduplication against an accumulated corpus, selection of diverse few-shot examples, and dynamic adjustment of category weights. It also provides specific software versions, hardware configurations, and pipeline parameters used for reproducibility, highlighting the engineering challenges of creating diverse and unique datasets at scale.
  48. Create a LangChain Deep Agents Harness Profile for NVIDIA Nemotron 3 Ultra to Improve Performance | NVIDIA Technical BlogThis post details the process of creating a LangChain Deep Agents harness profile for NVIDIA Nemotron 3 Ultra to improve agent accuracy without fine-tuning. It introduces the concept of harness engineering as an alternative to model fine-tuning, focusing on adapting the agent's interaction with the model. The post outlines a methodology involving running evaluation benchmarks, analyzing failures, proposing harness profile changes (e.g., adding middleware like ReadFileContinuationNoticeMiddleware to guide tool usage), and re-validating the improvements. It also discusses automating this harness engineering loop using agentic proposers and managed services like LangSmith Engine.
  49. NVIDIA Vera CPU Boosts AI Factory Throughput to Accelerate Agentic Workloads | NVIDIA Technical BlogThis post details how the NVIDIA Vera CPU architecture addresses bottlenecks in agentic AI workloads by providing significantly improved sustained per-core performance under full socket load. It highlights Vera's 1.8x faster cores for RL training throughput, its monolithic design with 88 Olympus cores for 40% lower peak loaded latency and over 3x per-core memory bandwidth at half the power of traditional x86 CPUs for agentic inference, and its ability to minimize CPU-side stalls and KV-cache evictions to maximize GPU utilization and session efficiency.
  50. Building an Analysis AI Agent for Industrial Alarm Management with NVIDIA Nemotron | NVIDIA Technical BlogThis post details the construction of a per-alarm analysis AI agent for industrial alarm management using NVIDIA Nemotron open models and the NVIDIA OpenShell secure runtime. It describes the agent's workflow: gathering context from structured (SQL, data warehouses) and unstructured (playbooks, manuals) sources using NVIDIA NeMo Retriever and Apache Vanna, performing specialist metric checks (e.g., anomaly detection, Fourier analysis) with GPU-accelerated libraries (cuDF, cuVS, cuFFT, cuML), and iteratively synthesizing observations, root-cause hypotheses, remedies, and recommended actions. The agent is exposed via a single HTTP endpoint for seamless workflow integration and leverages agent orchestration and sandboxed execution via OpenShell for enterprise deployment. The post highlights the use of Nemotron 3 Nano for orchestration and Nemotron 3 Super for complex reasoning, with optimized NIM containers for low-latency inference.
  51. Mastering Agentic Techniques: AI Agent Reinforcement Learning | NVIDIA Technical BlogThis post details the practical application of reinforcement learning (RL), specifically RLVR and GRPO, for customizing language models and agents in domain-specific enterprise workflows. It introduces NVIDIA's Nemotron 3 Super and the NeMo RL ecosystem (NeMo Gym, NeMo Data Designer) as tools for open model post-training, scalable environment-based evaluation, verifiable reward design, and synthetic data generation. The post outlines a minimal RL loop, emphasizing the importance of starting with evaluation, defining clear task requirements, designing trustworthy reward functions, and conducting iterative, small-scale experiments with continuous logging. It also contrasts RL with other techniques like RAG, prompting, and SFT, providing guidance on when to use each based on the available signal and desired behavior. The core contribution is a practical guide to implementing verifiable RL training for agents, highlighting the role of environments and verifiers in scoring success and updating model weights.
  52. How to Govern Autonomous Agents in Enterprise AI Factories | NVIDIA Technical BlogThis post introduces the NVIDIA Secure Agent Workspace Reference Design, a system for governing autonomous agents in enterprise AI factories. It details a two-phase implementation: perimeter security (provisioning managed workspaces, enforcing SSO, network lockdown, human approval for actions, centralized logging) and runtime security inside the VM (active sandboxing, signed security policies, credential protection via proxies, continuous verification). It also outlines the setup of agent blueprints for repeatable workflows and deployment strategies on Red Hat OpenShift Virtualization and Microsoft Azure.
  53. Deploy a Production-Ready NVIDIA AI-Q Blueprint on Oracle Cloud Infrastructure | NVIDIA Technical BlogThis post details the deployment of the NVIDIA AI-Q Blueprint, an open-source reference for long-horizon AI agents, on Oracle Cloud Infrastructure (OCI). It provides a concrete example of how to provision OCI resources (OKE, Load Balancer, Vault) using Terraform and deploy the AI-Q workloads (backend, frontend, PostgreSQL) using Helm. It outlines the necessary prerequisites, architecture overview of AI-Q and its OCI deployment, and step-by-step instructions for setting up the infrastructure and installing the AI-Q endpoint. This contributes a practical, cloud-native deployment pattern for complex agentic AI systems.
  54. Q&A: How KRAFTON Built PUBG Ally, a Co-Playable Character Powered by NVIDIA ACE | NVIDIA Technical BlogThis post details the technical architecture of KRAFTON's PUBG Ally, a co-playable character powered by NVIDIA ACE. It highlights the on-device pipeline combining NVIDIA ASR, a 2B-parameter Mistral-NeMo-Minitron-2B SLM, and a custom TTS model. Key engineering efforts focused on minimizing latency through prompt engineering for KV cache optimization and a System 1/System 2 architectural split, where a behavior tree handles fast, reactive gameplay and the SLM handles deliberate reasoning. The post also discusses domain adaptation for PUBG-specific terminology and grounding responses in actual game state by constraining the operational environment (single map, mode, fixed item taxonomy). The quantized SLM's ability to run on GPUs with as little as 8GB VRAM is also noted.
  55. Build an AI Scientist for Life Science Discovery with NVIDIA BioNeMo Agent Toolkit | NVIDIA Technical BlogThis post details how the NVIDIA BioNeMo Agent Toolkit enables AI scientists to perform biomolecular tasks by providing accelerated, agent-ready interfaces (BioNeMo Skills). These skills package core capabilities like structure prediction, molecular generation, docking, sequence analysis, and genomics as callable services accessible through hosted or local NIM deployments. The post emphasizes that BioNeMo Skills and Model Context Protocol server wrappers document model purposes, input requirements, expected artifacts, and failure modes, allowing agents to autonomously discover, select, invoke, and interpret biomolecular models. Empirical benchmarking demonstrated that integrating BioNeMo Skills doubled agents' token efficiency and increased task completion rates from 57.1% to 100%, transforming agent workflows from isolated model calls into iterative, production-ready research loops.
  56. How Telcos Build Autonomous Networks with Agentic AI | NVIDIA Technical BlogThis post details how telcos can build autonomous networks using agentic AI. It introduces a mental model for problem-solution loops in telecom operations (execute, optimize, discovery paths) and outlines the key building blocks of a telco autonomy platform. This platform includes telecom-domain models (Nemotron, NV-Tesseract), agent harnesses (NVIDIA Agent Toolkit), and secure runtimes (OpenShell). It emphasizes the use of NVIDIA technologies like NeMo Data Designer and Safe Synthesizer for data generation, and AI-Q for deep research skills. The post highlights practical applications in autonomous anomaly detection and remediation in SR-MPLS networks and AI-driven wireless network algorithm discovery.
  57. Enable Real-Time AI for High-Speed Data Acquisition with DAQIRI | NVIDIA Technical BlogIntroduces NVIDIA DAQIRI (Data Acquisition for Integrated Real-time Instruments), a high-performance networking library within the NVIDIA Holoscan Platform. DAQIRI enables direct, zero-copy streaming of high-bandwidth detector data to GPU memory, bypassing the Linux kernel via DPDK and GPUDirect. It facilitates real-time AI processing, filtering, compression, and adaptive control for scientific instruments and industrial applications. The post details its architecture, key features (high throughput, low latency, zero memory copy to GPU, YAML-driven configuration, C++/Python APIs), and provides a walkthrough of its configuration. It highlights its application in the A-GHOST project at CERN for real-time analysis of HL-LHC data.
  58. Building AI Agents for AR Glasses and XR Devices with NVIDIA XR AI | NVIDIA Technical BlogThis post introduces NVIDIA XR AI, an open-source library for building intelligent agents for AR glasses and XR devices. It details the modular architecture, which separates media transport, model services (NVIDIA Cosmos, Nemotron), enterprise connectivity (MCP), and agent orchestration (NeMo Agent Toolkit). The post provides a step-by-step guide to building a first agent, including starting AI services, running a sensor-first agent, connecting enterprise data via MCP, and adding agent orchestration.
  59. Build On-Device AI Companions with the NVIDIA ACE Game Agent SDK and Unreal Engine 5 Plugins | NVIDIA Technical BlogThis post introduces the NVIDIA ACE Game Agent SDK Beta and associated Unreal Engine 5 plugins, extending agentic AI infrastructure to game development. It details the SDK's Agent, Chat, and RAG APIs for building on-device AI companions and NPCs, enabling dynamic reasoning, conversation, and tool-assisted actions. The post also highlights new ACE Unreal Engine 5 plugins for ASR (nemo-conformer-ctc-120m), SLM (Qwen 3.5 4B with GGUF support), and TTS (Chatterbox Turbo 350M), all optimized for local, low-latency, RTX-accelerated workflows. Examples from PUBG: BATTLEGROUNDS and Total War: PHARAOH demonstrate the practical application of these tools for creating interactive game characters.
  60. NVIDIA Achieves Leading Agentic Coding Performance on First Agentic AI Benchmark | NVIDIA Technical BlogThis post introduces the Artificial Analysis AgentPerf (AA-AgentPerf) benchmark, the first open, multi-vendor benchmark for measuring concurrent AI agent support under real-world coding trajectories. It details the benchmark's methodology, which captures agentic workload complexity (non-deterministic sequences, tool call latencies, variable sequence lengths) using private, representative test sets. The post highlights NVIDIA GB300 NVL72's performance, demonstrating up to 20x higher concurrent agent throughput per megawatt compared to NVIDIA H200, attributed to optimizations like WideEP/DeepEP, DeepGEMM, fused MoE, and NVLink scale-up. It also projects future performance gains with the NVIDIA Vera Rubin platform.
  61. Deploy Long-Context Reasoning and Agentic Workflows with MiniMax M3 on NVIDIA Accelerated Infrastructure | NVIDIA Technical BlogThis post details the deployment of MiniMax M3, a 428B parameter multimodal MoE model with 1M-token context, on NVIDIA Blackwell infrastructure. It highlights MiniMax Sparse Attention (MSA) for efficient KV cache access and reduced per-token compute cost. The post provides deployment instructions for TensorRT LLM, SGLang, and vLLM, showcasing performance insights on Blackwell Ultra with up to 10x higher throughput than H100. It also covers scaling with NVIDIA Dynamo for distributed inference and customization with NVIDIA NeMo Framework for fine-tuning (SFT, LoRA) and reinforcement learning, including context parallelism up to 128k tokens.
  62. Delivering Lifecycle Control for AI Infrastructure at Scale with NVIDIA DGX Spark Enterprise Manageability | NVIDIA Technical BlogThis post introduces NVIDIA DGX Spark and GB10 systems with a new Enterprise Manageability framework. This framework provides a modular operational stack for AI infrastructure, supporting agentless SSH execution with standardized JSON outputs for integration into existing IT tools (CMDB, SIEM, monitoring). It covers six lifecycle phases: procurement, provisioning, monitoring, maintenance, incident response, and end-of-life. Key technical contributions include tools like `spark_diagctl.py` for L1 health checks and L2 diagnostics bundles, `reset_reason_reporter.py` for root cause analysis of reboots, and `spark_updatectl.py` for coordinated multilayer update management with staged rollouts and rollback capabilities. The custom installation process leverages cloud-init and OEM data partitions for preconfiguration and supports air-gapped deployments. Security features include verified boot integrity checks, encryption-at-rest state reporting, and APT signing verification.
  63. Accelerating Federated Learning Research with AI Agents and NVIDIA FLARE Auto-FL | NVIDIA Technical BlogThis post introduces NVIDIA FLARE Auto-FL, an automated research loop for federated learning that leverages AI agents and structured experimentation. It details the system's components, including a control plane (`program.md`), fixed benchmark contracts, an experiment ledger (`results.tsv`), and literature-grounded recovery mechanisms. The contribution lies in operationalizing agent-led coding into a controlled experiment workflow, enabling researchers to evaluate more FL strategies efficiently and reproducibly by constraining mutation surfaces and ensuring comparable metrics.
  64. Evaluate Clinical ASR Models Faster with Agent Skills and NVIDIA Nemotron Speech | NVIDIA Technical BlogThis post details a workflow for evaluating clinical ASR models using agent skills, NeMo Data Designer, and Nemotron Speech. It introduces a process for generating pronunciation-aware synthetic audio, reviewing clinical terms, and benchmarking ASR performance. The workflow leverages agent skills to guide the process, starting with defining clinical profiles and iteratively improving the benchmark based on ASR evaluation results. The core technical contribution is the pipeline for creating pronunciation-validated synthetic audio data, which addresses the challenges of rare clinical terminology and privacy concerns associated with real patient data.
  65. NVIDIA Nemotron 3 Ultra Powers Faster, More Efficient Reasoning for Long-Running Agents | NVIDIA Technical BlogIntroduces NVIDIA Nemotron 3 Ultra, a 550B-parameter Mixture-of-Experts model designed for frontier reasoning and orchestration in long-running agent workflows. Key technical contributions include: hybrid Mamba-Transformer layers for efficient long-context handling, NVFP4 quantization for cross-architecture GPU deployment with up to 5x higher throughput, LatentMoE for efficient expert routing, and multi-token prediction for improved generative speed. The model also incorporates Multi-Teacher On-Policy Distillation (MOPD) for continuous improvement and domain specialization, trained on a large dataset with open recipes and weights. Performance benchmarks show Nemotron 3 Ultra delivering leading accuracy with higher throughput and lower cost compared to other open models.
  66. Build Personal AI Agents on Windows PCs with New Tools from Microsoft and NVIDIA | NVIDIA Technical BlogThis post details the introduction of secure agent sandboxing on native Windows using Microsoft eXecution Containers (MXC) and NVIDIA OpenShell. It highlights performance improvements for agentic inference through optimizations in llama.cpp (Multi-Token Prediction, Programmatic Dependent Launch) and vLLM, achieving up to 2.6x throughput gains. The post also introduces multi-GPU support for RTX PCs via tensor parallelism in llama.cpp and LM Studio, and updates to agent capabilities with NemoClaw, Hermes Agent, and H Company's Holo 3.1 models, focusing on easier setup, native app integration, and GPU performance.
  67. Deploy Self-Evolving Agents for Faster, More Secure Research with a Hermes Agent and NVIDIA NemoClaw | NVIDIA Technical BlogThis post details the deployment of self-evolving AI agents using Hermes Agent and NVIDIA NemoClaw. It showcases how to securely integrate internal data (Outlook, Slack) with public data (GitHub, NVIDIA developer forums) using the OpenShell runtime, which enforces network policies and brokers credentials. A key contribution is the demonstration of agent self-improvement, where agents learn new skills (e.g., formatting daily digests) and persist them across deployments via snapshots, enabling faster and more secure research.
  68. Deploy Agentic-Ready AI at the Edge with Memory Efficiency in NVIDIA JetPack 7.2 | NVIDIA Technical BlogJetPack 7.2 introduces one-command deployment of NVIDIA NemoClaw on Jetson devices, enabling agentic AI applications. It adds NVIDIA agent skills for Jetson (device-side and BSP-side) to automate development tasks like Linux customization, memory optimization, and model benchmarking. JetPack 7.2 also brings Multi-Instance GPU (MIG) support to Jetson Thor for isolated GPU partitioning, and Super Mode for Jetson AGX Orin 32 GB for higher AI performance. Official Yocto Project support is also included.
  69. Run Local AI Agents with Faster Models and Multi-Node Clustering on NVIDIA DGX Spark | NVIDIA Technical BlogThis post details the introduction of NVIDIA DGX Spark with NVIDIA NemoClaw for local AI agent execution. It covers a streamlined installation process for DGX Spark's June 2026 release, integrating NemoClaw, OpenShell runtime, and agent harnesses like OpenClaw. It highlights the ability to run agents locally with enhanced security and privacy, and introduces multi-node clustering capabilities using NVIDIA Sync to connect multiple DGX Spark units for increased memory and throughput, supporting larger models and multi-agent pipelines. Performance improvements for Qwen3.6-35B with vLLM and NVFP4 quantized checkpoints are also presented.
  70. Advancing AI Infrastructure for Agentic AI with NVIDIA DOCA In-Silicon Security | NVIDIA Technical BlogThis post details how NVIDIA BlueField DPUs and DOCA security capabilities provide in-silicon security for agentic AI factories. It introduces DOCA Argus for runtime threat detection via memory analysis on the DPU, DOCA Vault for zero-trust file access, and DOCA Flow for high-speed network policy enforcement. The key technical contribution is the shift to a hardware-enforced, workload-independent security layer on the DPU, operating independently from the host system to protect AI infrastructure, workloads, agents, and data at AI scale, even if the host is compromised. It highlights the speed of DPU-based detection (1000x faster than agentless) and enforcement (800 Gb/s).
  71. NVIDIA DSX OS Delivers Open, Modular Software for Operating AI Factories at Scale | NVIDIA Technical BlogIntroduces NVIDIA DSX OS, an open-source, modular software suite for operating AI factories at scale. Key contributions include DSX Exchange for IT/OT communication and agentic interfaces, DSX MaxLPS and DSX Flex for power and efficiency optimization, NVIDIA Infra Controller (NICo) and BlueField DPUs for provisioning and tenant isolation, and NVSentinel/Fleet Intelligence for health monitoring and automated remediation. This post details how these components enable faster time to revenue, better efficiency (up to 40% more GPUs), and higher reliability for AI factories.
  72. DynoSim: Simulating the Pareto Frontier | NVIDIA Technical BlogIntroduced DynoSim, a workload-driven discrete-event simulation of the NVIDIA Dynamo serving stack. DynoSim combines measured engine forward-pass timing, scheduler cores, router and planner behavior, and KV cache effects on a virtual timeline to rapidly explore the Pareto frontier of LLM serving configurations. It achieves high simulation speed (e.g., 1500x real time) by composing components as actors on a discrete-event timeline, enabling the screening of thousands of configurations before hardware validation. The simulation architecture includes a replay harness, single-engine simulation with scheduler fidelity (e.g., vLLM, SGLang) informed by AIConfigurator (AIC) for timing, and multi-engine simulation modeling system behaviors like routing and distributed caching. Experiments demonstrate the benefits of KV-aware routing over round-robin for improved prefix cache reuse and reduced TTFT.
  73. Run Step 3.7 Flash on NVIDIA GPUs with Enterprise-Ready Multimodal AI | NVIDIA Technical BlogThis post details the deployment of Step 3.7 Flash, a 198B-parameter Mixture-of-Experts vision-language model, on NVIDIA-accelerated infrastructure. It covers deployment options using open-source frameworks (SGLang, TensorRT-LLM, vLLM), production-ready deployment via NVIDIA NIM as containerized inference microservices with standardized APIs, and Day 0 fine-tuning with the NVIDIA NeMo framework (including Automodel and Megatron-Bridge recipes). The post also mentions the use of NVFP4-quantized checkpoints for boosted inference and the availability of GPU-accelerated endpoints for prototyping and evaluation.
  74. Automating and Optimizing Financial Signal Discovery with Multi-Agent Systems | NVIDIA Technical BlogThis post details the application of agentic AI systems, orchestrated by the NVIDIA NeMo Agent Toolkit, for automating financial signal discovery. It describes a system of three specialized agents (Signal Agent, Code Agent, Evaluation Agent) that work in a continuous loop to identify, code, and backtest financial signals. The post highlights the use of Nemotron models for hypothesis generation and tool-calling, the structured library of mathematical operators provided to the Signal Agent, and the config-driven workflow management enabled by the NeMo Agent Toolkit for iterative refinement and optimization of signals based on metrics like Rank IC.
  75. Mastering Agentic Techniques: AI Agent Customization | NVIDIA Technical BlogThis post details nine techniques for customizing AI agents to excel at specialized tasks beyond general-purpose foundation models. It covers prompt engineering, retrieval-augmented generation (RAG), agent tool and skill injection, supervised fine-tuning (SFT), parameter-efficient fine-tuning (PEFT), Direct Preference Optimization (DPO), and reinforcement learning with verifiable rewards (RLVR). Each technique is explained with its mechanism, use cases, and limitations, providing guidance on selecting the appropriate method based on the need for better information, instructions, or more reliable behavior.
  76. Add a Specialized Deep Research Skill to Agent Harnesses | NVIDIA Technical BlogThis post introduces the AI-Q skill, a portable agent skill that packages the deep research pipeline of the NVIDIA AI-Q Blueprint. This skill enables agent harnesses (like Claude Code, Codex, and LangChain Deep Agents) to delegate deep research tasks, including multi-document synthesis and decision briefs backed by enterprise data, to a local or hosted AI-Q server. The skill manages intent classification, clarification, shallow and deep research, and evaluation, returning a structured report with citations. It addresses the complexity of grounding agents in enterprise data, routing queries, managing authentication, and preserving source attribution. The post also details secure MCP integration patterns for AI-Q as an MCP client, enabling it to connect to authenticated MCP servers as data sources using various authentication scenarios (no per-user auth, backend/app credentials, or forwarding user's bearer token). Deployment properties for regulated industries, such as data locality, self-hostable models, and auditability, are emphasized.
  77. NVIDIA-Verified Agent Skills Provide Capability Governance for AI Agents | NVIDIA Technical BlogThis post introduces NVIDIA-verified agent skills as a mechanism for capability governance in AI agents. It details the verification process, which includes cataloging, scanning for software and agent-native risks using SkillSpector, cryptographic signing for authenticity, and documentation via machine-readable skill cards. The skill card schema is described, outlining information such as skill function, ownership, licensing, dependencies, and known limitations. The post also highlights how verified skills extend AI governance to the capability layer, complementing existing runtime guardrails like NVIDIA NeMo Guardrails.
  78. Mastering Agentic Techniques: AI Agent Evaluation | NVIDIA Technical BlogThis post introduces a shift in evaluating AI agents from static model benchmarks to dynamic, trajectory-aware metrics. It details five practical tips for evaluating AI agents as production systems: prioritizing task success rate (TSR) over accuracy, evaluating full trajectories (plans, tool calls, reasoning, outcomes) rather than just final answers, making tool usage a first-class signal (measuring selection precision, recall, and schema compliance), scoring reasoning quality and efficiency (soundness, tokens, steps, latency), and building transparent, customizable evaluation from day one. The post emphasizes that reliable agentic systems require tracking outcomes, tool usage, reasoning, and cost together, and wiring those signals into the development process.