AI Lab
Official NVIDIA announcements — model releases, product launches and research, each one summarized with every source covering it, by AIBriefs. RSS
NVIDIA's post evaluates Tensor Core methods that decompose FP32 operands into lower-precision components and recombine results in FP32, boosting matmul throughput while keeping accuracy near native FP32. The approach needs no retraining or changes to model weights or architecture, addressing why many MLIPs still run inference in FP32.
The open-weight 30B MoE (~4B active) transcribes 40 languages, translates 31 to and from English, and separates up to eight speakers. It ranks first on Hugging Face's Open ASR leaderboard for English at 5.33% average word error rate, but handles only recorded audio, not live streams.
CUDA 13.4 enforces a transparent per-workload memory boundary in a new time-sharing mode, so allocations above a set size fail as they would on a smaller GPU. Workloads can request an exact amount or a guaranteed minimum with a growable maximum, and pick sm-sharing compute mode.
NVIDIA built Nemotron-Personas-France with Pleias and Nemotron-Personas-Belgium with Pleias and KU Leuven, part of an open Nemotron data program covering locale-grounded persona datasets for 12 countries. The blog walks through post-training open Nemotron models so agents reflect regional regulations and cultural nuance.
C-BRAIN, a consortium of 18 institutions launched in November 2025 and led by Dr. Randall Bateman, is working with NVIDIA on open-source AI tools to mine previously inaccessible Alzheimer's data, including unpublished findings and negative results. It targets 57 million people living with dementia and 10 million new cases a year.
Profiling Ling-3.0-flash at batch size 1 on DGX Spark showed decoding was memory-bound, so NVIDIA applied Humming MXFP4 MoE kernels, online FP8 quantization of the LM head, and NEXTN multi-token speculative decoding. The model has 124B total parameters with 5.1B active per token.
The autonomous recovery engine, previously limited to Slurm-based HPC workloads, now covers Kubernetes workloads managed with NVIDIA Run:ai. It detects, diagnoses and recovers from hardware, software and network anomalies across the stack so multi-node training jobs resume automatically.
The Global Performance Tuner, released in TensorRT 11.1, does automatic, accuracy-aware, end-to-end tuning by treating alternative scheduling strategies as searchable build routes. A two-part NVIDIA series covers the motivation and math, then the algorithm, architecture, and best practices.
8B total / 4B effective parameters with 64K context, post-trained from Google's Gemma 4 E4B using reinforcement learning only, no supervised fine-tuning. Runs in 3.1–4.3 GB VRAM on RTX, Jetson and DGX Spark; NVIDIA-measured top score on 15 of 21 benchmarks, but Tau2-Bench trails Qwen3.5-4B 55.1 to 75.1.
NVIDIA's walkthrough uses Hermes and Nemotron 3.5 Bolt on RTX Spark to turn a prompt into a project plan, with Astra building assets and Isaac Lab training the navigation policy. Physics comes from Newton and simulated visuals from Omniverse RTX libraries, with policies evaluated locally in Isaac Sim.
Firebird secured a U.S. export license for 45,000 NVIDIA Grace Blackwell GPUs, with phase one under construction alongside Kazakhtelecom at Data Center Valley for a Q3 2027 commercial launch. A planned second phase would bring Firebird's total potential capacity in Kazakhstan to roughly 500 megawatts.
NVIDIA's KvikIO Python API sits above the cuFile C library to move files directly from storage into GPU memory, skipping CPU staging. The post cites the 2025 Stanford AI Index: model compute doubles roughly every five months, LLM training datasets every eight.
Eight linked DGX Spark nodes pool 1 TB of unified memory, up from 128 GB on a single unit, enough to run Z.ai's full GLM 5.3 (753B) locally. GLM 5.3 Flash hit roughly 80 tokens per second across the cluster; NVIDIA Sync's Cluster Assistant support for eight nodes arrives by year's end.
Minerva correlates telemetry across GPUs, SuperNICs, switches, storage and schedulers into a dynamic knowledge graph, analyzing billions of data points per hour to trace slow training jobs to the culprit component. It ships with a native AI assistant grounded in live telemetry and connects to external agentic tools.
Revolut is rebuilding its stack on a single foundation model, PRAGMA, running on Blackwell Ultra GPUs and cuDF/Dynamo-Triton, serving an initial 80 million customers and engineered to scale 3x. Starling Bank trained an agentic model on Civo's U.K. sovereign cloud; Lloyds built transaction model ATUM on RTX PRO 6000 Blackwell GPUs.
Cosmos ranks top five on six open-model leaderboards. Five cloud-native Data Factory blueprints ship via GitHub, Microsoft Azure, Nebius and CoreWeave, covering AV dataset generation, defect image generation and video augmentation. Porsche is building a shared AV data flywheel across Volkswagen Group brands on it.
NVIDIA says power is now infrastructure's biggest bottleneck, with over $300 billion in equipment backlogged and 200+ gigawatts of projects stuck in U.S. interconnection queues. DSX Sim, Flex, MaxLPS and Exchange tie grid, facility and compute planning together; GE Vernova, Hitachi Energy and Schneider Electric are building on it.
Shopify post-trained NVIDIA Nemotron 3.5 Lightning for its Sidekick GraphQL subagent, beating a leading frontier model on its internal benchmark while cutting serving costs 90% and responding 7x faster. Zooplus hit 86% search accuracy in an AWS prototype using Nemotron Retriever.
StormScope EU is a regional short-term forecasting model trained on European weather data, predicting cloud cover across 16 Meteosat channels at 10-minute intervals from EUMETSAT satellite observations. NVIDIA demoed a live pipeline with Earthmover at GTC Berlin; AXA is using FourCastNet 3 to study regional cyclones.
deepset is expanding its Haystack Enterprise Platform, the foundation of the European Commission's AI@EC shared platform, adding cost comparison, answer-quality evaluation and problem tracing. Deloitte is building agentic healthcare and employment systems for Poland on open models.
NVIDIA's Kumo Relational reportedly outperforms supervised machine learning on relational data with no training required, per a Hugging Face blog post. The post frames it as a training-free approach to relational-data prediction.
Hugging Face post walks through a sim-to-real pipeline that trains the Microduck robot in Nvidia's Isaac Lab 3.0 before deploying the policy on physical hardware.
NVIDIA's Kumo Forecast combines pretrained time-series representations with historical pattern retrieval and interpretability, per a Hugging Face blog post. The post frames retrieval of past patterns as the mechanism layered on top of the pretrained representations.
RFD4-Proteina, a flow-matching model from the Baker Lab and NVIDIA, co-designs protein sequences and all-atom structures, skipping the separate inverse-folding handoff to models like ProteinMPNN. It ships as an open BioNeMo model plus a NIM microservice, with a tutorial targeting PD-L1 binders via reward-guided inference-time search.
NVIDIA Omniverse published 2,000+ simulation-ready 3D assets on the Hugging Face Hub, open source for robotics simulations and AI factory digital twins. The library is searchable in plain English, and Hugging Face added OpenUSD support to host the assets.
A5X instances pair Vera Rubin NVL72 with ConnectX-9 SuperNICs and Google Cloud's Virgo Network; GA lands next month with Europe expansion planned. Google Cloud plans up to 960,000 Rubin GPUs across multiple sites.
FOX is on GitHub in early access for building always-on factory manager agents plus specialists for visual inspection, root-cause analysis and fleet management. Siemens is integrating FOX-built agents into Intelligence Center X for an estimated 20% cost savings, and cut inspection calibration time 40% at its Erlangen factory.
NVIDIA's CUDA-X library moves large sparse-system solves onto GPUs; Dassault Systèmes reports 8-10x speedups in Abaqus and Applied Materials up to 10x in Ginestra versus CPU-only. ZEISS SMT ran an EUV lithography plasma benchmark 4x faster, finishing in six days and 14 hours.
Nebius, Nscale and Verda are adopting Vera CPUs to handle the data processing and code execution behind agentic workloads, alongside GPUs running models. Bull, Dell, HPE, Lenovo and Supermicro are building Vera-based systems; ClickHouse, SAP and Elastic reported strong early results on analytics and semantic search.
The reference design combines OpenShell, an open-source secure runtime that traces agent actions and enforces policy, with Sentry, which runs on BlueField-4 DPUs in a hardware trust domain outside the agent's reach. Perplexity ran 9 models with root access across 108 attempts to break out of a VM; none breached the boundary.
New Agent Knowledge Foundation bundles CUDA-X libraries (cuDF, cuGraph, cuML, cuOpt, cuVS) plus an Auto-Ontology library that maps raw company data to business concepts. Ships alongside Nemotron Retriever and the new Kumo family of open models, OpenShell secure runtime and NemoClaw blueprints.
NVIDIA added Nemotron 3.5 Super VL and Nemotron 3.5 Bolt, plus Nemotron Retriever embedding/reranking/parsing models, the Kumo family for structured business data, Cosmos 3.5 world models, Alpamayo 2 Edge for autonomous vehicles, BioNeMo and Ising for quantum. Nemotron was built with the Nemotron Coalition, including European partners Black Forest Labs, H-Company and Mistral.
The DSX Blueprint, part of NVIDIA DSX Sim, ships reference digital twin assets, sample scripts and workflows built on OpenUSD and Omniverse libraries. Agents assemble and revise OpenUSD scenes from site plans and SimReady assets so developers can test layouts before engineering review.
DSX POD packs seven Vera Rubin racks, networking and liquid cooling into a shipping container, cutting deployment at a prepared site from six months to as little as 14 days. Partners ship units in early 2027; DSX MaxLPS software claims up to 30% less power for the same token output.
Jensen Huang's Berlin keynote followed a weeklong European road trip (Paris, London, Stockholm, Dublin) with Nvidia committing nearly $10 billion across the region and working with 12,000+ partners. He pitched AI as a five-layer stack and unveiled DSX POD, modular AI factories built, cabled and tested before delivery.
NVIDIA's Cosmos3-H-Dreams model is presented as a generative simulation system for clinical surgery, per a Hugging Face blog post. The post frames it as bringing surgical scenarios to life through synthetic video rather than real footage.
Barcelona Supercomputing Center's AI Factory Accelerator Programme is now in full operation, with some services accessible via the NVIDIA Inception portal. The first AI factory in Spain has delivered more than 800 services to private and public-sector clients since its April 2025 launch.
Chancellor Friedrich Merz and Jensen Huang discussed industrial AI at Berlin's Kraftwerk during the Invest in Germany Summit, where the federal government signed a Joint Declaration of Intent with Nvidia on physical AI models. Nvidia also announced planned investments in four German startups.
Jane Street ML engineers implemented Mamba-2's forward pass in NVIDIA's cuTile, a Python DSL for tile-level GPU programming that targets Tensor Cores automatically. The post walks through the three custom kernels behind the forward pass and the cuTile optimizations used.
The European Broadcasting Union, whose 115+ member organizations reach over 1 billion people across 55+ countries, is expanding its sovereign framework with NVIDIA to add sovereign cloud providers, open model developers and integrators. Microsoft Sovereign Cloud (Azure Local, Foundry Local) is among the providers mapping the standards to deployment options.
RTX Spark pairs a Blackwell RTX GPU with a Grace CPU for up to 128GB unified memory and 1 petaflop of AI compute; laptops from ASUS, Dell, HP, Lenovo, Microsoft and MSI start at $2,399. Artificial Analysis clocked a local agent workload on Qwen3.8 27B at 18.7 minutes.
A 4B-parameter BF16 model built on a looped (latent-recurrent) normalized-GPT architecture with interleaved sliding-window and full-attention layers. Supports up to 256K tokens of context, English only, positioned for general-purpose reasoning.
Hugging Face post details an Indic-language pretraining run at 2.8 trillion tokens, paired with an open data generation pipeline. The pipeline is released openly rather than kept proprietary.
Muse Spark Open is an 816B mixture-of-experts model with 42B active parameters and a 1M token context window, scoring 39 on the Artificial Analysis Intelligence Index — the #1 US open-weights model. Meta trained it on NVIDIA GB300 NVL72 for agentic coding, and it adds audio and video perception to image and document understanding.
Nemotron 3.5 Super VL is an open 120B-parameter model with 12B active, up to 1M tokens of context, and text, image, document and video understanding. NVIDIA says it matches its largest reasoning model at roughly one-fifth the parameter count and runs 5x faster than the next best in its class.
At the AI in Science Summit in Dublin, the European Commission launched the AI in Science Enablers Initiative, tied to RAISE, with NVIDIA committing accelerated tools, compute, cloud resources and training. NVIDIA cites 250+ open models, 250 open datasets and 1,400+ open source projects on Hugging Face and GitHub.
Siemens' Fuse EDA AI Agent used NVIDIA NeMo Gym to verify Liberty (.lib) files against known references, raising its benchmark composite score from roughly 55% to ~95% on standard-cell characterization tasks. The fixes came from revising the agent's skill and MCP tool descriptions, which also cut turns and output tokens.
The 74B-parameter, 9B-active MoE model ships in 24GB via 2-bit quantization with quantization-aware training to retain quality. NVIDIA says it is purpose-built for agentic AI on RTX Spark laptops, which arrive Oct. 16.
The five-year compute commitment accompanies a UK deployment push spanning Fuse Energy's customer AI on B300 Blackwell Ultra for 450,000+ households and OneAdvanced's fine-tuned LLM for NHS primary-care triage. NVIDIA also joins the Earthshot Prize as a Global Alliance Founding Partner.
Cities and partners across Belgium, Finland, France, Italy, the Netherlands, Poland, Sweden and the U.K. are building physical AI agents and digital twins on NVIDIA accelerated computing. Cosmos and Metropolis VSS models interpret video for traffic, equipment and safety issues; Nemotron powers reasoning agents.
NVIDIA's cuDF skills for Apache Spark use agents to rewrite user-defined functions into GPU-accelerated equivalents in minutes, reporting 22x+ speedups in Spark jobs. The post covers testing, benchmarking, and integration of the rewritten UDFs.
Nvidia uploaded a GR00T-N1.7 checkpoint to Hugging Face, tagged AppleToPlate-cosine40k-checkpoint-15000, suggesting a robot manipulation policy trained on an apple-to-plate task for 40k steps with a cosine schedule.
The release ships as part of the NVIDIA DSX AI Factory Platform, with source in the DSX AI Factory GitHub org and four independently adoptable components: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management, and Workload Map. The community edition carries no NVIDIA enterprise support.
NVIDIA's GTC Berlin post highlights French companies deploying agentic computing across space, financial services, the public sector and scientific research. It credits the February 2025 AI Action Summit and the 2025 and 2026 Choose France Summits with anchoring major compute commitments on French soil.
Altair-1, the first of a $1 billion, 50-satellite constellation, reached orbit flying NVIDIA Jetson AGX Orin modules running Mistral models for autonomous Earth observation. Twake.ai, built on NVIDIA Nemotron, became the sole sovereign digital workplace for 2,500+ French public organizations.
ReplaySSM stores SSM state as a checkpoint plus a buffer of recent inputs, separating output computation from state updates. NVIDIA says SSM layers can bottleneck latency up to 100K tokens of context at serving batch sizes. It is integrated into TensorRT-LLM, vLLM, and FlashInfer, and runs in production in GLM-5.3-Flash and Kimi-K3.
NVIDIA contributed Vera Rubin rack and compute tray designs to the Open Compute Project via its open MGX architecture, announced at the OCP Global Summit in San Jose. Partners also published AI factory design and grid resilience standards, and NVIDIA and Cisco unveiled USFP, the industry's first 400Gb/s module.
Nvidia's FP4 QAD checkpoint weighs 102GB versus 133GB for its NVFP4 build, per r/LocalLLaMA. Red Hat's NVFP4 version keeps MoE experts in FP4 and the rest in BF16, runs on vLLM, and accepts text, image, and video input.
Nvidia's GLM-5.3-Flash-FP4-Local upload has drawn 108 likes and 2,942 downloads since its Oct 8 posting. The FP4 quantization targets local deployment of the GLM-5.3-Flash model.
NVIDIA's Omniverse blog details developers directing frontier models like GPT-6 Astra via natural language to wire up Omniverse libraries for physics (ovphysx), rendering (ovrtx) and scene updates (ovstage). Projects include a warehouse humanoid simulator and an autonomous-driving test environment based on San Francisco's Market Street.