Agentic AI Infrastructure Acceleration with BlueField DPUs

8/11/2026

This 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.

8/11/2026

Introduces NVIDIA Nemotron 3.5 Lightning, a 30B parameter MoE model with 3B active parameters, optimized for high-volume, low-latency execution in always-on AI agents. Details its features like speculative decoding, harness-optimized training, and quantization (NVFP4 and BF16 checkpoints) for strong accuracy and up to 4x output speed. Also introduces NVIDIA NeMo Switchyard for intelligent model routing to optimize task allocation across different models. The post highlights Nemotron 3.5 Lightning's performance on the Artificial Analysis Intelligence Index and PinchBench, demonstrating its efficiency for agentic tasks.

8/11/2026

This post introduces NVIDIA NeMo Switchyard, an orchestration layer for agentic AI that enables dynamic routing of workloads across specialized and frontier models. It details how NeMo Switchyard addresses the challenge of selecting the optimal model for each task by considering model capabilities, cost profiles, and infrastructure signals. The post describes the provider-agnostic SDK (libsy), the separation of routing logic from specific model providers, and the architecture that supports stateless or stateful routing. It also outlines different routing algorithms, including tuning-free approaches like LLM classifiers, stage routers, and escalation routers, and discusses the importance of infrastructure for seamless handoffs and observability.

8/10/2026

This 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.

8/4/2026

Introduced 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.

7/30/2026

This 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.

7/30/2026

This 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.

7/29/2026

This post details the technical implementation of self-hosting a validated AI coding assistant. It covers deploying StarCoder2-7B as an NVIDIA NIM with an OpenAI-compatible endpoint, integrating this endpoint into IDEs via custom configurations, and implementing NVIDIA NeMo Guardrails as a proxy for policy enforcement. The post outlines the architecture, prerequisites, and step-by-step instructions for setting up the NIM, configuring the IDE, and deploying NeMo Guardrails with specific YAML configurations for input flows and prompts. It also touches upon the CI verification gate and metrics loop for enhanced security and traceability.

7/27/2026

Introduces NVIDIA Ising Calibration 1.5, a 31B-parameter vision language model for diagnosing and tuning quantum processors. Highlights its NVFP4-quantized version for single GPU or NVIDIA DGX Spark deployment, an 11.4% reduction in model size at BF16 precision, and training on diverse qubit modalities. Details its evaluation using the QCalEval benchmark, showcasing state-of-the-art zero-shot and in-context learning performance. Mentions availability of full-parameter checkpoints, quantized versions, open datasets, and deployment blueprints 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.

7/27/2026

This post introduces the NVIDIA Labs Object-Oriented Agents (NOOA) framework, an open-source research preview designed to improve agent performance. NOOA structures agents as single Python classes, treating methods as capabilities, fields as state, and docstrings as prompts. It leverages type annotations as enforced contracts and uses an LLM-driven loop to complete method bodies marked by ellipses at runtime. The framework implements six key model-facing interface ideas: typed input/output, pass by reference, code as action, programmable loop engineering, explicit object state, and model-callable harness APIs. A novel long-term memory subsystem is also introduced, allowing agents to curate typed, relational memory in a human-readable SQLite store, supporting knowledge accumulation and efficient context management through pass-by-reference, 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 costs compared to prior harnesses, with reproducible, inspectable workflows and open evaluation methodology.

7/27/2026

Introduced NVIDIA Nemotron 3 Ultra, a 550B parameter Mixture-of-Experts Hybrid Mamba-Attention model with a 1M-token context length, specifically trained on a synthetic RTL dataset. Demonstrated its effectiveness in agentic RTL coding workflows when combined with the ACE-RTL agent, achieving a 97.1% average pass rate on the CVDP benchmark and outperforming other models in accuracy and token efficiency. Detailed the synthetic data generation pipeline used for training Nemotron 3 Ultra on RTL coding tasks.

7/24/2026

Introduces ModelExpress (MX), a system for accelerating the distribution of model artifacts. MX prioritizes direct GPU-to-GPU P2P RDMA transfers via NIXL, bypassing object storage and host memory. It optimizes cold starts by streaming from object storage without landing on disk, using multithreaded tensor readers and pipelining reads with GPU placement. For clusters with shared disk caches, MX uses a Model Cache Service to coordinate downloads, ensuring the fleet populates the cache only once. When GPUDirect Storage (GDS) is supported, MX reads checkpoint files directly from local storage into GPU memory through NIXL's multithreaded GDS backend, bypassing host memory. MX also supports loading local checkpoints through ModelStreamer, overlapping disk I/O with GPU placement. For subsequent workers, MX treats serving replicas as live weight sources, transferring tensors directly from source GPU to target GPU over P2P RDMA via NIXL. The control plane discovers peers through Redis, Kubernetes CRDs, or k8s-service metadata. MX also optimizes NIXL memory registration overhead.

7/21/2026

This 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.

7/21/2026

This post introduces the NVIDIA Rubin GPU architecture and the Vera Rubin platform, detailing hardware advancements specifically designed to 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 (22 TB/s bandwidth), NVLink 6 (3,600 GB/s), PCIe Gen 6, and Confidential Computing with TEE-I/O. 2. MoE optimization: Enhanced Tensor Memory Accelerator with inline descriptor updates for TMA, reducing metadata-management and data-movement overhead for MoE models. 3. GEMM acceleration: Doubled Tensor Core throughput along the K-dimension, enabling fewer iterations for GEMMs and improving efficiency for both context and decode operations at high tensor-parallel scale. 4. Vera Rubin NVL72 platform: Integration of liquid cooling, DSX MaxLPS power smoothing, cable-free MGX architecture, and hot-swappable NVLink switch trays for rack-scale deployment of multitrillion-parameter models.

7/21/2026

This post introduces the NVIDIA Vera CPU and its Olympus core, specifically designed to maximize single-thread performance for agentic AI workloads. It details the architectural innovations of the Olympus core, including its optimized branch predictors, wide instruction throughput, deep out-of-order execution, and advanced cache subsystem with a graph prefetcher. The post also highlights NVIDIA Spatial Multithreading for flexible resource partitioning and the Scalable Coherency Fabric for high on-die bandwidth, along with SOCAMM2 memory modules for increased bandwidth. Secure scale-up connectivity through NVLink-C2C, PCIe 6.4, CXL 3.1, and Confidential Computing is also discussed.

7/20/2026

This 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.

7/20/2026

This 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.

7/16/2026

This post introduces NVIDIA NemoClaw as a framework for building autonomous agents that integrate context-aware video AI into enterprise workflows. It details how NemoClaw, in conjunction with NVIDIA Metropolis Blueprint for Video Search and Summarization (VSS) and NVIDIA AI Blueprint for Retrieval-Augmented Generation (RAG), enables agents to ingest video, capture user intent via human-in-the-loop prompts, retrieve organizational context from documents, and generate structured reports with recommended actions. The post demonstrates this by creating a 'healthy eating coach' agent that analyzes food videos, assesses eating habits against nutritional guidelines, and automatically creates Jira tickets for follow-up actions. The technical contribution lies in the orchestration of these components to move from video analysis to programmatic action within enterprise systems.

7/16/2026

This post details the system-level capabilities of the NVIDIA BlueField-4 DPU and Vera BlueField-4 STX Storage Processor, co-designed with DOCA software, to address the infrastructure demands of agentic AI. It highlights how these components offload, accelerate, and isolate networking, storage, security, and context management services by integrating high-speed connectivity, embedded compute, local memory, PCIe Gen6, and inline acceleration. Specific DOCA services like Host-Based Networking (HBN), ASTRA, Memos for KV cache reuse, and security services are described as enabling programmable infrastructure services for AI factories, improving GPU utilization, reducing inference latency, and enhancing multi-tenant isolation.

7/14/2026

This post details five practical lessons learned from the NVIDIA Nemotron Model Reasoning Challenge on Kaggle, which involved over 5,000 participants aiming to improve AI reasoning. The key technical contributions and observations include: 1. The importance of making chain-of-thought data verifiable by auditing intermediate steps and repairing flawed traces, rather than just adding data. 2. Designing reasoning to fit within token budgets by compressing repeated structures and preserving logic to leave room for further reasoning. 3. Separating stable knowledge (reusable patterns, lookup tables) from live reasoning to avoid asking the model to rediscover structure repeatedly. 4. The effectiveness of using tools to generate and audit high-quality training data, focusing on the quality of reasoning traces. 5. The significance of measuring performance by task type and validating against real failure modes, with participants leveraging community discussions for technique sharing and workflow optimization on specific hardware (Google Cloud G4 VMs with NVIDIA RTX PRO 6000 Blackwell GPUs).

7/10/2026

Introduces 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.

7/9/2026

This 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.

7/8/2026

This post details the creation of 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, using evaluation benchmarks and agent harness profiles as customization entry points. A specific example demonstrates how to fix a `read_file` tool failure by introducing `ReadFileContinuationNoticeMiddleware` to the harness profile, improving test results and overall evaluation scores. The post also discusses automating this harness profile creation process using an iterative loop similar to agentic workflows.

7/7/2026

This post introduces an AI agent for industrial alarm management built using NVIDIA NeMo Agent Toolkit, Nemotron open models, and OpenShell secure runtime. The agent automates evidence gathering, specialist analysis, and action recommendation for industrial alarms. It integrates with GPU-accelerated libraries like cuDF, cuVS, cuFFT, and cuML for data handling and analysis. The agent uses Nemotron 3 Nano for orchestration and Nemotron 3 Super for reasoning, with optimized NIM containers for low-latency inference. It employs tools for accelerated retrieval of structured information (Text-2-SQL with Apache Vanna) and accelerated RAG with NVIDIA NeMo Retriever for unstructured data. Specialist checks include anomaly detection using NVIDIA nv-tesseract and OCR. The agent's output is a structured evidence package (observation, root-cause hypothesis, remedy, recommended action) and supporting trace, delivered via a single HTTP endpoint.

7/1/2026

This post details the application of Reinforcement Learning (RL), specifically Reinforcement Learning with Verifiable Rewards (RLVR) and Group Relative Policy Optimization (GRPO), for customizing language models and agents in domain-specific enterprise workflows. It introduces NVIDIA's Nemotron 3 Super and 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 the minimal RL loop (policy model, task, action, environment, verifier, rollouts, policy update) and provides guidance on choosing between RAG, prompting, SFT, DPO, RLHF, and RLVR based on the available signal and desired behavior. It highlights GRPO as a practical default for RLVR tasks and discusses the importance of environment-based RL for agentic systems that fail across long-horizon workflows. The post also covers synthetic data generation using NeMo Data Designer and the use of NeMo Gym for generating scored trajectories.

6/29/2026

This 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.

6/26/2026

This 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 outlines the use of Terraform for provisioning OCI resources (VCN, OKE cluster, Load Balancer, Vault) and Helm for installing the AI-Q backend, frontend, and PostgreSQL workloads on OKE. The post covers prerequisites, architecture overviews of both the AI-Q multi-agent system and its OCI deployment, and step-by-step instructions for configuration, infrastructure creation, and Helm chart installation, including secret management for API keys.

6/25/2026

This post details KRAFTON's implementation of PUBG Ally, a co-playable character powered by NVIDIA ACE. It highlights the technical architecture involving on-device automatic speech recognition, a 2B-parameter small language model (Mistral-NeMo-Minitron-2B), and text-to-speech. Key engineering efforts focused on minimizing latency through on-device deployment, prompt engineering for efficient KV cache usage, and architecting a System 1 (behavior tree) / System 2 (language model) pipeline to handle fast, reactive gameplay separately from deliberate reasoning. Domain adaptation for PUBG-specific terminology and grounding responses in game state were also addressed.

6/23/2026

This post introduces the NVIDIA BioNeMo Agent Toolkit, which enables AI scientists to perform biomolecular tasks by providing accelerated, agent-ready interfaces (BioNeMo Skills). These skills package core capabilities such as structure prediction, molecular generation, docking, sequence analysis, and genomics as callable services accessible through hosted or local NIM deployments. 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 with high efficiency and reliability. 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.

6/23/2026

This post introduces a mental model for agents to move through problem-solution loops in telecom operations and outlines the key building blocks of a telco autonomy platform. It details the role of data and models (NeMo Data Designer, NeMo Safe Synthesizer, Nemotron, NV-Tesseract), agent harnesses (Agent Toolkit), and secure execution runtimes (OpenShell) in enabling autonomous agents for tasks like anomaly detection and remediation, and AI-driven wireless network algorithm discovery. It also discusses the use of digital twins and specialized skills for deep research.

6/22/2026

Introduces 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.

6/17/2026

Introduces NVIDIA XR AI, an open-source library for building intelligent agents for AR glasses and XR devices. This framework enables real-time visual and voice interaction, enterprise data access, and tool integration by connecting XR devices to GPU-accelerated AI services. The architecture is modular, separating components like media transport, model services (including NVIDIA Cosmos for vision and Nemotron for language), enterprise connectivity via Model Context Protocol (MCP), and agent orchestration using frameworks such as NVIDIA NeMo Agent Toolkit, supporting flexible deployment and multi-user scenarios. Developers can quickly prototype XR agents by cloning the public beta repository, running sample agents, integrating enterprise data through MCP servers, adding agent orchestration, and optionally incorporating CloudXR for rendered spatial experiences.

6/16/2026

Introduced the NVIDIA ACE Game Agent SDK Beta and associated Unreal Engine 5 plugins for building on-device AI companions and NPCs. The SDK provides a C/C++ agentic framework with Agent, Chat, and RAG APIs, optimized for small models and RTX hardware. The Unreal Engine plugins offer local ASR, SLM, and TTS models with Blueprint and C++ support, enabling low-latency AI character interactions within games.

6/12/2026

This post introduces the Artificial Analysis 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 for capturing agentic workload complexity, including non-deterministic sequences, tool call latencies, and variable sequence lengths, using private, representative test sets. The post highlights NVIDIA GB300 NVL72's leading performance on this benchmark, 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 Vera Rubin platform.

6/12/2026

This post details the deployment of the MiniMax M3 model, a 428B parameter Mixture-of-Experts model with 1M-token context and native multimodality, on NVIDIA Blackwell accelerated infrastructure. It highlights the MiniMax Sparse Attention (MSA) mechanism, which replaces standard quadratic attention with a pre-filtering stage for efficient long-context reasoning and agentic workflows, achieving significant speedups in prefill and decoding. The post also outlines deployment strategies using open-source inference engines (TensorRT LLM, SGLang, vLLM), large-scale serving with NVIDIA Dynamo, and customization with the NVIDIA NeMo Framework, including support for N-D parallelism and context parallelism up to 128k tokens.

6/9/2026

This 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.

6/9/2026

This 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.

6/9/2026

This 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.

6/4/2026

Introduces NVIDIA Nemotron 3 Ultra, a 550B-parameter Mixture-of-Experts model designed for agent orchestration and long-running agent workflows. Highlights architectural innovations including hybrid Mamba-Transformer layers, NVFP4 quantization for cross-architecture GPU deployment (achieving up to 5x higher throughput), LatentMoE for efficient expert routing, and multi-token prediction for improved generative speed. Details the Multi-Teacher On-Policy Distillation (MOPD) training method for continuous improvement and domain specialization, and outlines expanded training data including domain-specific pre-training data (synthetic legal, Wiki-based, refreshed GitHub tokens) and post-training data (SFT samples, RL tasks, RL environments).

6/2/2026

This post introduces Microsoft eXecution Containers (MXC) and NVIDIA OpenShell for secure, on-device agent execution on Windows PCs, enabling turnkey agent sandboxing. It details performance improvements in llama.cpp (2x for Qwen 3.5/3.6 27B, 1.6x for Qwen 3.5/3.6 35B MoE) via Multi-Token Prediction and Programmatic Dependent Launch, and in vLLM (2.6x improvement) with optimizations for MoE models and CUDA Graphs. It also highlights multi-GPU support in llama.cpp and ComfyUI through tensor parallelism, offering up to ~2x memory capacity and ~1.8x compute performance on RTX PCs. Updates to NVIDIA NemoClaw, Hermes Agent, and H Company's Holo 3.1 models are also discussed, focusing on expanded agent capabilities, ease of setup, native app integration, and performance on NVIDIA GPUs.

6/2/2026

This post introduces an open-source example of deploying self-evolving AI agents using Hermes Agent and NVIDIA NemoClaw. It details the architecture, which combines a reasoning model (NVIDIA Nemotron 3 Super), a harness (Hermes Agent), and a secure runtime (NVIDIA OpenShell). The post provides a step-by-step guide for installation, configuration, and teaching the agent new skills, emphasizing the ability to learn and persist these skills across conversations and deployments. It also highlights the security aspects of mixing public and private data through OpenShell's network policy enforcement and credential brokering, and introduces observability features using Agent Trajectory Format (ATIF) traces with NeMo Relay and Arize Phoenix.

6/2/2026

JetPack 7.2 introduces agentic AI capabilities to the edge with Jetson devices. Key contributions include: one-command deployment of NVIDIA NemoClaw for secure, agentic AI applications; NVIDIA agent skills for Jetson (device-side and BSP-side) to automate development tasks like Linux customization, memory optimization, and model benchmarking; and Multi-Instance GPU (MIG) support on Jetson Thor for partitioned, deterministic execution of mixed-criticality workloads. Official Yocto Project support and Super Mode for Jetson AGX Orin 32 GB are also included.

6/2/2026

This post introduces NVIDIA NemoClaw, a new open-source blueprint for running local AI agents on DGX Spark. It details a streamlined installation process that bundles open models, an agent harness (OpenClaw), and the NVIDIA OpenShell runtime. The post highlights performance improvements for Qwen3.6-35B with NVFP4 quantization and MTP optimizations, achieving up to 2.6x faster inference. It also introduces the cluster assistant in NVIDIA Sync for setting up multi-node DGX Spark clusters (2-4 nodes) to provide unified memory for larger models and distributed agent pipelines, simplifying ConnectX-7 networking configuration.

6/1/2026

This post introduces the integration of NVIDIA DOCA in-silicon security capabilities with NVIDIA BlueField DPUs to secure AI factories for agentic AI. It details how BlueField DPUs provide a hardware-enforced, workload-independent security layer that operates independently from the host system. The post highlights the DOCA security stack, including DOCA Argus for runtime threat detection via memory analysis, DOCA Vault for zero-trust access, and DOCA Flow for high-speed network policy enforcement. It emphasizes that these capabilities operate at AI speed and scale, protecting AI infrastructure, workloads, agents, and data.

6/1/2026

Introduces 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.

5/30/2026

Introduced 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.

5/29/2026

This post details the deployment and fine-tuning of the Step 3.7 Flash multimodal AI model on NVIDIA GPUs. It highlights the model's architecture (198B parameters, MoE, 11B active parameters, 256k context window, native image/video input) and its suitability for agentic workflows. The post outlines deployment options using SGLang, TensorRT-LLM, and vLLM, and emphasizes production-ready deployment via NVIDIA NIM microservices with OpenAI-compatible APIs. It also describes Day 0 fine-tuning using the NVIDIA NeMo framework, including Automodel for SFT and LoRA, and NeMo Megatron-Bridge for large-scale training, demonstrating efficient fine-tuning on Hopper GPUs.

5/21/2026

This 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.

5/20/2026

This post details nine techniques for customizing AI agents to enhance their performance on specialized tasks. It covers prompt engineering, retrieval-augmented generation (RAG), agent tool and skill injection, supervised fine-tuning (SFT), and parameter-efficient fine-tuning (PEFT). It also introduces advanced methods like Direct Preference Optimization (DPO) and reinforcement learning with verifiable rewards (RLVR), often paired with Group Relative Policy Optimization (GRPO), for nuanced alignment and improved reasoning by leveraging preference signals or objective correctness criteria. The post provides criteria for selecting the right techniques based on use case needs for better information, instructions, or more reliable behavior.

5/20/2026

Introduced NVIDIA AI-Q as a specialized deep research skill for agent harnesses. This skill enables agents to delegate complex research tasks (multi-document synthesis, decision briefs, long-horizon analysis with source attribution) to a local or hosted AI-Q server, receiving structured reports with citations. The integration supports various MCP authentication scenarios (no per-user auth, backend/app credentials, forwarding bearer tokens) and allows AI-Q to run in diverse environments (developer laptop, Kubernetes, air-gapped data centers) while keeping sensitive data within controlled environments. The AI-Q pipeline is engineered for research quality, comprising intent classification, clarification, shallow research, deep research, and evaluation stages.

5/20/2026

Introduces NVIDIA-verified agent skills as a mechanism for capability governance, addressing the need for trust and transparency in deploying AI agent capabilities. Details the verification process, including cataloging, risk scanning (SkillSpector), cryptographic signing, and skill cards. SkillSpector checks for both conventional software risks and agent-specific risks like prompt injection and excessive agency. Cryptographic signing provides verifiable provenance and integrity checks for downloaded skills.

5/19/2026

This post introduces a refined approach to evaluating AI agents, distinguishing it from AI model evaluation. It emphasizes measuring end-to-end system performance through dynamic, trajectory-aware metrics rather than static benchmarks. Key contributions include defining five practical tips for agent evaluation: prioritizing task success rate over accuracy, evaluating full trajectories (plans, tool calls, reasoning, outcomes), making tool usage a first-class signal (selection precision, schema compliance), scoring reasoning quality and efficiency (soundness, tokens, latency), and building transparent, customizable evaluation systems from day one. It highlights the importance of tracking metrics like Task Success Rate (TSR), Tool Call Accuracy, and Trajectory Efficiency to ensure reliable agent behavior in production.