News
Updates, product notes and practical coverage for AI platforms and operators.
NVIDIA has released Nemotron 3.5 Super VL, an open and customisable multimodal mixture-of-experts model for agents working across text, images, documents and video. The 120-billion-parameter system activates 12 billion parameters per token. NVIDIA claims strong single-GPU throughput and highlights customer post-training gains, which adopters should reproduce on their own workloads.
NVIDIA has released Nemotron 3.5 Bolt 74B-A9B-24GB, a native two-bit multimodal model designed to fit within 24GB of memory. It combines a 74-billion-parameter, nine-billion-active architecture with a 256K context window, image and video inputs, tool calling and local vLLM deployment, targeting capable agents on a single workstation-class GPU.
NVIDIA says RTX Spark laptops are now shipping from six major PC makers, combining a Grace CPU, Blackwell RTX graphics, up to 128GB of unified memory and one petaflop of FP4 AI performance. The Windows systems target local models, creative workloads and agents including OpenClaw, with cloud connectivity available for larger tasks.
NVIDIA and Microsoft have introduced RTX Spark systems for local AI agents on Windows, with laptop preorders open and compact desktops arriving later. The platform combines Blackwell graphics, Grace processing, unified memory and CUDA, while a forthcoming DGX Station for Windows targets trillion-parameter enterprise workloads.
NVIDIA has announced Nemotron 3 Large Telco Model, a 30-billion-parameter open model fine-tuned by AdaptKey for network configuration and customer-incident workflows. The 6 October article also releases an end-to-end NeMo fine-tuning recipe. Operators gain more customisation and deployment control, but still need strong privacy engineering, evaluation and production governance.
NVIDIA has announced a 64GB DGX Spark configuration due from six manufacturer partners on 23 October. Its 2 October post pairs the smaller system with Sync Cluster Assistant for connecting two units, while a model launcher is planned later in the month. The hardware targets local agent development, but buyers should check model fit, costs and availability.
NVIDIA says CoreWeave has made Vera Rubin NVL72 systems available to early cloud customers, with Cognition running production workloads. Its 30 September account also details Vera CPU sandbox testing and a new Forge environment for improving agents. The claims are significant for AI infrastructure, but reported throughput gains need workload-specific validation before buyers extrapolate them.
NVIDIA has introduced an Open Agent Safety Platform combining its broadly available OpenShell runtime with a Sentry reference design on BlueField-4 hardware. OpenShell enforces agent permissions outside the model, while Sentry monitors from a separate trust domain. The launch offers layered controls, but organisations must still define policies, test escapes and verify real-world deployment claims.
NVIDIA, Google DeepMind and EMBL-EBI have released predicted structures for protein complexes from more than 2,800 viruses through the AlphaFold Database. NVIDIA says BioNeMo accelerated the work and is opening the prediction pipeline. The resource can help researchers form and prioritise hypotheses, but computational structures still require confidence checks and laboratory validation.
NVIDIA has released Isaac ROS 5.0 as a free, open-source toolkit for robotics developers. It adds agent-ready skills and documentation, ROS Lyrical and Ubuntu 24.04 support, and perception tools that run across Jetson devices. The release could shorten development work, but teams still need physical testing, safety controls and measured performance before deploying robots.
NVIDIA has launched DSX Ready, a qualification programme for products used with its AI-factory reference designs. It begins with battery energy storage systems and cooling distribution units, naming initial partners in both categories. Qualification is product-specific and does not replace site engineering, so buyers still need to test fit, resilience and local constraints.
NVIDIA says the ANAB has accredited its Halos AI Systems Inspection Lab as an ISO/IEC 17020 inspection body. Its physical-AI safety update spans vehicles and robots, connecting hardware, software, simulation and repeatable inspections. The accreditation covers the lab’s scoped inspection work; it is not blanket certification of every Halos-based robot or autonomous vehicle.
NVIDIA says its Vera Rubin NVL72 preview achieved up to 3.7 times GB300 NVL72 throughput on Qwen3-VL in MLPerf Inference v6.1. The same release reports strong Blackwell rack scaling and software gains. These are workload-specific benchmark results, not a universal production speed or cost guarantee.
NVIDIA, Google and Emerald AI have launched the AI Energy Management Alliance to develop performance-based rules for data centres that can adjust electricity demand. The initiative targets faster, more reliable grid connections, but it is an alliance and proposed framework rather than a new power connection or guaranteed reduction in bills.
Perplexity Portable Computer is expanding to Windows, bringing local AI task execution to compatible NVIDIA RTX systems. The announced configuration requires at least 24GB of graphics memory, making eligibility a critical detail. Local processing and optional cloud assistance offer different trade-offs, so users should examine the route taken by each task.
NVIDIA has expanded CUDA-Q with an open-source logical compilation framework for fault-tolerant quantum development. The release focuses on representing programmes, comparing architectures and tracing resource assumptions through compilation. Reported development-time savings should not be confused with faster quantum execution or proof that commercially useful fault-tolerant hardware has arrived.
NVIDIA has agreed to acquire Hugging Face for US$12.93 billion, promising the model hub will remain open, multi-cloud and multi-accelerator. The proposed deal would pair Hugging Face's developer community and catalogue with NVIDIA infrastructure, model evaluation, inference and deployment resources, while leaving developers free to choose hardware, models and services.
NVIDIA has introduced PAIR, an open-source router that distributes local AI inference across compatible PCs. The IFA 2026 update also covers faster llama.cpp and vLLM inference, streamlined local-agent setup, and RTX Spark Windows PCs planned for October. Teams should test routing, device membership and data boundaries before relying on distributed local inference.
NVIDIA has added NVHBM to its NVLink Fusion platform, putting the memory controller inside the HBM stack. The company says the design can improve bandwidth and power efficiency while giving cloud and AI-chip builders a more standard route to custom, rack-scale systems for enterprise AI teams.
NVIDIA says its Groq 3 LPX inference system is now in full production alongside Vera Rubin NVL72. The company positions the LPU-based platform for faster token generation and long-context responsiveness in agentic applications, citing early deployments at Nebius and CoreWeave and a tighter integration with its wider AI-factory stack.
Firebird has opened an NVIDIA-powered AI factory in Armenia that the companies describe as the CIS region’s largest. The project combines Dell systems, NVIDIA DSX infrastructure and planned Rubin and Blackwell GPU deployments, signalling a push to build AI capacity closer to regional research, enterprise and public-sector users.
Baseten has added NVIDIA Nemotron 3.5 ASR Streaming models to its model library, offering English and 40-locale multilingual transcription through NVIDIA NIM. Baseten reports support for 100 concurrent real-time WebSocket streams on one H100, with finalisation latency below 140 milliseconds in its test. Production results will vary by workload.
NVIDIA and roughly 40 technology organisations have formed the Open Secure AI Alliance to develop shared, open defences for AI agents and software. NVIDIA is contributing models, data and its new NOOA agent-harness research project, although delivery and governance details remain limited.
NVIDIA is deploying its Vera CPU in the electronic design automation workflows used to build future processors. Early tests on selected Cadence Jasper and Synopsys VCS workloads reached up to 1.5 times higher performance, but full benchmark and external availability details were not disclosed.
NVIDIA has open-sourced a GPU-accelerated Medical Physics Simulation framework within Isaac for Healthcare. It combines anatomy and device physics, simulated sensors and robot learning so developers can generate difficult clinical scenarios and evaluate medical robotics policies before costly hardware and laboratory testing.
NVIDIA has introduced Spectrum-6, a 102.4-terabit-per-second Ethernet switch system designed for Vera Rubin AI factories. Early deployments are planned by major cloud and infrastructure operators, but NVIDIA’s performance, efficiency and reliability figures remain vendor claims that buyers should validate against their workloads.
NVIDIA says its Vera Rubin rack-scale AI platform is entering full production across a global partner network. The company reports large gains in throughput per megawatt and token cost, alongside new deployments by cloud and model providers, though independent workload testing is still needed.
NVIDIA has released the 4-billion-parameter Cosmos 3 Edge world model through Hugging Face. The open model combines visual reasoning, prediction and action generation for robots and vision agents, with real-time operation demonstrated on Jetson Thor and scripts for post-training specialised policies.
NVIDIA Agent Toolkit now includes open Omniverse libraries for sensor simulation, GPU-accelerated physics and validation of simulation-ready 3D assets. The company is also releasing a Blender blueprint, while SideFX, PTC and several startups begin integrating the components into existing design and simulation workflows.
Bristol Myers Squibb is deploying a second NVIDIA DGX SuperPOD built from eight Vera Rubin NVL72 systems. The pharmaceutical group plans to make the unified environment available across its research organisation for model training, predictions and agentic drug-discovery workflows using NVIDIA BioNeMo tooling.
Hugging Face and NVIDIA have integrated Diffusers-format image and video models with NeMo Automodel for distributed fine-tuning. The Apache 2.0 tooling supports direct Hub checkpoints, configurable parallelism and ready-made recipes for models including Wan, FLUX, HunyuanVideo and Qwen-Image.
Baseten has made NVIDIA's Nemotron 3 Embed 8B and 1B models available for dedicated inference. The pair targets enterprise and code retrieval with different balances of accuracy, throughput and indexing cost, giving developers a managed deployment option while published performance figures remain vendor-reported.