Knowledge catalogue
The Qwen3.8‑Flash‑Next 176 B multimodal MoE model uses a hybrid Gated DeltaNet / Qwen Sparse Attention (QSA) architecture that compresses long‑context history and aggregates tokens into micro‑blocks f
Qwen3.8‑Flash‑Next is a multimodal mixture‑of‑experts model with 125 B parameters that employs Gated DeltaNet (compressing historic context) and Qwen Sparse Attention (exact retrieval over full contex
COMPASS is a scalable cross‑embodiment robot navigation framework that builds on NVIDIA’s X‑Mobility pretrained policies, adding lightweight RL fine‑tuning specialists for new robot‑scene pairs to avo
NVIDIA’s NVLink Fusion integrates its custom NVHBM memory into the MGX rack‑scale architecture, enabling hyperscalers to deploy large numbers of custom XPUs with higher bandwidth, reduced silicon area
CUDA Python 1.0, released alongside CUDA 13.3, provides NVIDIA-maintained libraries and tools that expose the entire CUDA platform directly to Python through a single shared foundation. It adopts sema
Shadow Engine Recovery in NVIDIA Dynamo keeps a fully-initialized standby LLM engine on the same GPU, using the GPU Memory Service (GMS) to persist model weights and share them across engines without
NVIDIA’s Spectrum‑X Ethernet combines hardware‑accelerated adaptive routing, targeted congestion control and NIC‑based plane load balancing to meet the extreme bandwidth and low‑latency demands of gig
NVIDIA Groq 3 LPX, when paired with the Vera Rubin NVL72 platform, delivers a record‑setting 3,431 output tokens/second on the Artificial Analysis 100K context benchmark using Gemma 4 31B, demonstrati
NVIDIA’s new Scale‑In network infrastructure, the fifth pillar of its AI networking stack, is powered by BlueField‑4 DPUs, DoCA software, and Spectrum‑X Ethernet to deliver dedicated, host‑independent
NVIDIA’s new GPUs in the Vera Rubin and Blackwell families set a new benchmark for agentic‑AI performance per watt on AgentX, an open‑source inference test that models realistic multi‑step, tool‑using
Agentic AI workloads exhibit extreme variability—telemetry from over 163,000 sessions shows more than 97 % have unique workload trajectories, rendering traditional multi‑design CPU fleets ineffective.
AdaptGrow is a GPU‑accelerated matrix‑factorization framework that transforms rolling correlation or tail‑dependence matrices into hard clusters, soft factor loadings, and structural‑break signals for
NVIDIA DSX MaxLPS is a Dynamic Power Software (DPS) solution that reallocates power in real time across racks and GPUs to maximize performance per watt while operating within fixed facility limits. By
NVIDIA’s Agentic Variation Operators (AVO) architecture delivers a frontier‑level general‑purpose agent by integrating persistent memory, supervision, and tool use within the system harness rather tha
Recent frontier‑AI incidents show agents can bypass intended restrictions, underscoring the need to embed security controls at the runtime and infrastructure layers of an AI agent stack rather than wi
Generative recommenders (GRs) replace traditional embedding‑based methods with transformer‑style sequence models like HSTU and Semantic IDs, addressing scalability, cold‑start, and long‑tail issues in
NVIDIA FLARE orchestrates federated multimodal AI training by supporting both full‑model and parameter‑efficient (adapter‑based) communication patterns, using large‑object externalization, tensor stre
The blog describes the iterative development of a real‑time endoscopic tool‑segmentation application using NVIDIA Holoscan, HoloHub, and an AI coding agent. Each cycle consisted of engineering-defined
NVIDIA SkillEvaluator is an open‑source framework that measures how pre‑verified “skills” improve AI agent performance through static checks, distinctiveness analysis, and live task runs in isolated e
NVIDIA’s Cosmos 3 Edge is a 4‑billion‑parameter omni‑model (with a 2‑billion‑parameter Nemotron reasoner) that can run on an NVIDIA Jetson Thor, providing on‑device policy inference for robot manipula
NVIDIA’s ALCHEMI Toolkit delivers composable, PyTorch‑native GPU‑accelerated workflows for machine learning interatomic potentials (MLIPs), enabling AI coding agents to generate robust simulation code
NVIDIA cuML and cuVS now support multi‑GPU all‑neighbors kNN graph construction in UMAP, enabling end‑to‑end scaling of the dimensionality reduction algorithm to hundreds of billions of vectors. The i
The QAD (Quantization‑Aware Distillation) pipeline for Nemotron 3.5 Lightning leverages NVIDIA Model Optimizer to aggressively compress the model with W4A16/NVFP4 quantization while keeping accuracy c
A full‑stack observability framework for NVIDIA AI factories links telemetry from compute, networking, storage, orchestration and application layers using specialized tools (DCGM, NVSM, UFM, NetQ, NMX
Alibaba released the open‑weights Qwen3.8‑2.4T‑A95B (Qwen3.8‑Max), a fine‑grained mixture‑of‑experts model with 2.4 trillion parameters, hybrid full‑ and linear‑attention, a one‑million‑token context
NVIDIA JetPack 7.2.1 adds agentic video skills with the unified jetson‑videosdk, allowing programmable, device-aware video workflows that link developer intent to live device discovery and performance
NVIDIA Nemotron 3.5 Lightning is an open‑source mixture‑of‑experts language model totaling 30 B parameters with only 3 B active during inference, designed to serve high‑volume, low‑latency execution f
NVIDIA NeMo Switchyard is a routing platform that directs AI agent workloads to the most suitable specialized or frontier model for each step of a task, balancing performance, cost, and latency. It of
Meta released Muse Glimmer, a 30‑billion‑parameter dense language model with a context window exceeding 120 K tokens, designed for local, long‑running agentic AI workloads. The model is optimized to r
World Action Models (WAMs) use video-based world modeling instead of vision‑language backbones, giving robots a learned physics engine that supports zero‑shot transfer to new tasks, environments, and
NVIDIA Alpamayo 2 Super is a publicly available 34‑billion‑parameter vision–language–action model that merges a 32‑B Cosmos 3 Super Reasoner with a 2‑B action‑expert diffusion network. It produces uni
A combined architecture using KAI Scheduler and vCluster enables multiple teams to run fully isolated Kubernetes tenant clusters on a shared GPU node, each with its own control plane, RBAC, CRDs, and
NVIDIA’s Vera BlueField‑4 STX Storage Processor combines an 88‑core Armv9.2 design with spatial multithreading, a scalable coherency fabric and LPDDR5X memory to deliver high‑throughput storage proces
The article shows that dense‑attention performance in long‑context inference is governed by group size (query heads per KV head), head dimension, and sequence length, with prefill being compute‑bound
NVIDIA Video Codec SDK 13.1 adds AV1 hierarchical reference mode supporting up to 31 B‑frames and efficient iterative tuning that delivers significant bitrate savings in CQ and VBR modes. It enhances
NVIDIA’s AI Red Team found common failure modes in enterprise AI agents: weak access controls, unrestricted code execution via tools, unprotected network egress, and exposure of plaintext secrets. The
NVIDIA’s Exemplar Cloud study shows that identical H100‑based clusters can yield 8–12 % lower training throughput when configuration gaps—at the kernel, hypervisor, BIOS or NCCL levels—prevent reachin
NVIDIA **nvmath‑python v1.0** is a Python library that wraps CUDA‑X and NVPL math libraries (cuFFT, cuBLASLt, cuDSS, cuSPARSE, cuTENSOR, cuBLASMp) to give NumPy, CuPy, and PyTorch users GPU‑accelerate
Healthcare robotics struggles with data scarcity, limited generalization to rare clinical scenarios, and slow prototyping due to the need for annotated demonstrations and costly experimental setups. N
Applied Materials and NVIDIA have created an end‑to‑end digital development platform that unifies GPU‑accelerated tools (Ginestra, cuDSS, cuEST, PhysicsNeMo, Omniverso) with CUDA‑X libraries to accele
NVIDIA Ising Calibration 1.5 is a 31‑billion‑parameter vision‑language model that diagnoses and tunes quantum processors, delivering state‑of‑the‑art zero‑shot and in‑context learning on the QCalEval
NVIDIA’s Nemotron 3 Ultra, when paired with the ACE‑RTL agent, achieves a 97.1 % average pass rate on the CVDP benchmark across nine RTL task categories—surpassing GLM 5.2 and Kimi K2.6 while using up
The performance of AI agents depends not only on the underlying models but also heavily on their “harness”—the surrounding architecture that supplies context, state management, action execution, and t
NVIDIA ModelExpress (MX) is an agentic AI platform that streamlines the lifecycle of large‑model weights by automatically locating and using the fastest path to load them—prioritizing GPU‑to‑GPU P2P R
The NVIDIA OptiX Toolkit (OTK) is a BSD 3‑clause licensed collection of utilities that provides unified, minimal‑macro error checking across OptiX, CUDA runtime, and CUDA driver APIs, generating diagn
Prime Intellect Lab offers a streamlined, hosted reinforcement‑learning workflow that lets users customize the NVIDIA Nemotron 3 Nano in minutes. The process establishes a baseline, trains the model o
NVIDIA TensorRT’s IProgressMonitor API lets developers track and cancel long‑running engine builds in both Python and C++. By overriding `phase_start`, `step_complete`, and `phase_finish`, the monitor
NVIDIA’s Rubin GPU, the core of the Vera Rubin platform, delivers up to 10× the agentic inference throughput per watt compared with previous generations, using 336 billion transistors, 224 SMs, 896 Te
NVIDIA’s Vera CPU, built around the Olympus core, is engineered for agentic‑AI workloads that rely heavily on single‑thread performance, deep memory‑level parallelism, and efficient handling of irregu
The NVIDIA GB300 NVL72 platform set a world record by delivering 1,648 TFLOPs per GPU during the pre‑training of DeepSeek‑V3 671B, a mixture‑of‑experts (MoE) model. This achievement leveraged fifth‑ge
NVIDIA’s Omniverse libraries, now part of the NVIDIA Agent Toolkit, expose a modular C/Python SDK called **ovrtx** that enables real‑time, physically‑grounded sensor simulation (camera, LiDAR, radar)
NVIDIA’s sixth‑generation NVLink is a purpose‑built scale‑up network that delivers up to 3.6 TB/s per GPU and 260 TB/s rack‑level bandwidth with 130 TFLOPS in‑network compute—outperforming off‑the‑she
NVIDIA DeepStream 9.1 introduces AutoMagicCalib (AMC) and Multi‑View 3D Tracking (MV3DT) to automate camera calibration and maintain consistent 3‑D object IDs across multiple calibrated cameras, reduc
CUDA 13.3 introduces **clmad**, a hardware‑accelerated carryless multiply‑accumulate instruction available on all NVIDIA Ampere and newer GPUs (SM 80+), filling a long‑standing gap in GPU‑native binar
nanousd‑labs, part of NVIDIA Omniverse Labs, uses AI agents to generate lightweight, spec‑compliant USD runtimes directly from the USD Core Specification, sidestepping the need to adapt large legacy c
Autonomous coding agents such as Codex (GPT‑5.5) can fully automate reinforcement‑learning research workflows by provisioning GPU‑hosted environments, orchestrating experiments, and iteratively optimi
The NVIDIA Nemotron Model Reasoning Challenge on Kaggle attracted over 5,000 participants who all began from the same open model, benchmark, and infrastructure. The strongest solutions treated reasoni
NVIDIA Cosmos 3 was post‑trained in under a day using TAO agent skills and LoRA adapters, raising accuracy on the Woven Traffic Safety video QA dataset from 54.41 % to 93.35 %. The mixture‑of‑transfor
Guided diffusion‑based climate emulators (e.g., NVIDIA cBottle) steer generative models toward low‑probability weather states and use an odds‑ratio diagnostic—requiring second‑order derivatives—to rew
NVIDIA’s Ising Decoder demonstrates a 347.7‑fold improvement in logical error rate (LER) and a 7.3‑fold faster runtime over the Chromobius decoder for the d = 31 triangular color code at a physical er