AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
NVIDIA Developer Blog
First-party releases and research from NVIDIA Developer Blog. Headlines and excerpts link to the original articles.
Latest stories
Newest first
NVIDIA published an open post-training workflow for adapting Nemotron models to local rules and culture, releasing locale-grounded persona datasets for France and Belgium and a French public-service example built on Nemotron 3.5 Lightning.

NVIDIA's Factory Operations Blueprint coordinates multiple AI agents to investigate production problems such as failed circuit boards or rush orders, pulling evidence from separate factory systems. A Siemens report cited by NVIDIA estimates unplanned downtime costs an average large plant in the surveyed sectors $253 million a year.

NVIDIA detailed three optimizations for mixture-of-experts inference on DGX Spark that raised decode throughput by 42.5% and cut mean end-to-end latency by 24.5% on GSM8K, with no measurable accuracy change. The post also warns that combining the techniques requires a specific LM-head sharing interface, or performance drops below baseline.

NVIDIA Mission Control 2.5 extends its autonomous recovery engine beyond Slurm to Kubernetes workloads managed by NVIDIA Run:ai, letting faulty nodes be isolated and jobs restarted without waiting for root-cause diagnosis. The release also adds beta AI network anomaly detection, central Keycloak authentication, hardening updates and early access for Vera Rubin NVL72.

NVIDIA's TensorRT 11.1 Global Performance Tuner treats scheduling choices as searchable build routes over more than 200 compiler knobs, tuning 1.14x to 1.63x faster across ten evaluated workloads versus the default route. Accuracy acts as a filter and performance as the ranking criterion.

NVIDIA introduced the Global Performance Tuner in TensorRT 11.1, a feature that automatically searches over customizable build routes to find the fastest inference engine that still meets end-to-end accuracy targets.

NVIDIA published a walkthrough of an agent-assisted pipeline that plans, builds and trains a simulated obstacle course for the quadruped robot Microduck, using Hermes and Nemotron 3.5 Bolt for planning, Astra for 3D assets, and Isaac Lab for training on 4,096 parallel environments.

NVIDIA's KvikIO library gives Python and C++ users a higher-level interface over the cuFile API to read files directly into GPU memory via GPUDirect Storage, but reaching cuFile does not guarantee the accelerated path was used. The tutorial shows how to confirm it with cuFile statistics and how KvikIO and cuFile settings shape each read.

NVIDIA announced at GTC Berlin official support for eight-node DGX Spark clusters, scaling unified memory to 1TB so models like the full 753B GLM 5.3 checkpoint can be served locally, with the eight-node capability in NVIDIA Sync arriving by year's end.

NVIDIA released RFD4-Proteina, a BioNeMo open model co-developed with the Institute for Protein Design at the University of Washington, which co-designs protein sequences and all-atom structures and adds reward-guided inference-time search. At 200 residues it generates 3.8x faster and uses 84% less peak reserved memory than RFD3.

NVIDIA Sentry is a DOCA-based security framework running on BlueField-4 DPUs that monitors AI agents and enforces access policies from a hardware-isolated trust domain, with planned availability early next year.

NVIDIA published a developer guide to its Omniverse DSX Blueprint, a reference toolkit of digital twin assets, sample scripts and workflows for building AI factory campus models on OpenUSD, with frontier AI agents generating and revising geometry inside them.

NVIDIA published DSX POD, a standardized reference design for factory-built, container-based AI factories aligned with its Vera Rubin rack-scale platform. Partners integrate and validate compute, networking, storage, power and liquid cooling in containers before shipment, while site teams prepare foundations and utilities in parallel.

Jane Street machine learning engineers implemented Mamba-2's forward pass in NVIDIA's cuTile, a Python tile-level GPU DSL, and worked with NVIDIA to optimize three custom kernels. The post walks through the decay-cumsum, state-jump and fused output kernels plus compiler hints used to reach Tensor Cores.

NVIDIA RTX Spark pairs the GB10 system-on-a-chip with up to 128GB of unified memory to let developers run coding agents, fine-tune models, and compile code on a single Windows or WSL PC. NVIDIA claims 4.6x faster coding-agent performance, 2.7x faster fine-tuning, and 1.8x faster code compilation than Intel Panther Lake 12Xe.

Meta has released Muse Spark Open, an 816B-parameter mixture-of-experts coding model with 42B active parameters and a 1M-token context window, trained on NVIDIA's GB300 NVL72 platform for agentic coding. NVIDIA reports it sustains over 50K tokens per second per GPU and 600 tokens per second per user on that hardware.

NVIDIA released Nemotron 3.5 Super VL, a 120B-parameter open multimodal mixture-of-experts model that activates 12B parameters per token and is built for specialized enterprise agents. It combines text, image, document and video understanding with tool use, and can be post-trained with NVIDIA NeMo.

Siemens used NVIDIA's open NeMo Gym evaluation library to refine its Fuse EDA AI Agent for standard-cell characterization, raising a composite verification score from roughly 55% to 95% while cutting agent turns from 155 to 53 and output tokens from about 81K to 29K.

NVIDIA released open-source cuDF skills for Apache Spark that let coding agents rewrite CPU user-defined functions into GPU-accelerable forms, with one production rewrite cutting operator time from 274.55 to 12.78 hours across a 600-billion-row workload.

NVIDIA has open sourced the full Run:ai platform under its DSX AI Factory initiative, including four independently adoptable components: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management, and Workload Map. Eight DSX ecosystem partners are integrating it, and the integrated stack is available via GitHub and NGC as a Community Product without a paid subscription or enterprise support entitlement.