NVIDIA AI

What if your computer-use agent could learn a new command-line tool—operating it safely, without writing files or free-typing shell commands? With NVIDIA Nemotron open models, you can quickly train specialized agents for any CLI, using synthetic data and reinforcement learning—no usage logs required. Curious how it can be adapted for your own proprietary CLI environment? Join us live for a developer-focused livestream. In this interactive livestream, we’ll cover: - A live demo: Generating synthetic training data using NeMo Data Designer, validating examples, and fine-tuning with techniques like LoRA and reinforcement learning with verifiable rewards (RLVR) using NeMo RL. - Showcase how safety is built into every layer, from human-in-the-loop command approval to runtime isolation. - Explain why synthetic data generation and RL accelerate agent specialization without compromising accuracy or trust. Bring your questions about the workflow, model customization, and synthetic data generation, we’ll be answering them live throughout the livestream.

Join us live from NeurIPS for an inside look at groundbreaking research built with NVIDIA Nemotron, a family of open models with open weights, training data, and recipes. During the livestream, you’ll hear directly from NVIDIA researchers advancing the frontiers of AI: An introduction and insights from their NeurIPS-accepted research paper shaping new directions in machine learning The motivations and challenges behind their work Real-time answers to audience questions via live chat How these discoveries can be applied, extended, and integrated into future AI systems Whether you’re attending NeurIPS on-site or tuning in from anywhere, this interactive session is your chance to learn, engage, and connect with the minds driving the next wave of AI innovation.

"What makes it all connect together is NVLink” — Ian Buck, NVIDIA VP of Hyperscale and HPC. Built with 72 Blackwell GPUs linked by 130 TB/s of NVLink, GB200 NVL72 functions as one giant GPU—perfectly synchronized to power complex Mixture‑of‑Experts models like DeepSeek‑R1. GB200 NVL72 spreads experts across all 72 GPUs, communicating faster than ever, driving higher efficiency and performance. The result: up to 10× faster, 10× more efficient, and 10× more revenue per token than H200. Its extreme co‑design across compute, networking, and software - from chip to rack to data center - unleashes the full potential of Mixture‑of‑Experts computing. ➡️ Watch the full #SC25 special address and explore key announcements: https://nvda.ws/4oRd4qB

🎉 Introducing Nemotron-Elastic-12B 🎉 We're thrilled to announce the release of Nemotron-Elastic-12B, enabling high-quality reasoning capabilities across multiple model sizes with unprecedented efficiency! ✨ What Makes It Special: 🧠 Many-in-One Model, Single Training – Extract 6B, 9B, 12B, nested models from one training run using zero-shot slicing. No retraining. No separate checkpoints. ⚡ Constant Training Cost – Unlike traditional compression that scales linearly with the number of target sizes, Nemotron-Elastic maintains approximately constant training overhead regardless of how many variants you extract. For our 6B, 9B, and 12B family: 7.2× token savings vs. traditional methods. 💾 Constant Deployment Memory – Deploy all model variants (6B, 9B, 12B) using memory equivalent to storing just the largest model. All three Nemotron-Elastic variants fit in 24GB—a 2.25X reduction compared to maintaining separate checkpoints for each size. 🧮 Exceptional Reasoning – Built on a hybrid Mamba-2 + Transformer architecture, delivering on-par or better accuracy compared to models of the same size on reasoning benchmarks MATH-500, AIME, GPQA, LCB, etc.. 🎯 Ideal for Edge Deployment – Dynamic model selection on resource-constrained devices, from mobile to edge servers, without managing multiple checkpoints or retraining. 📚 Dive Deeper: 📖 Read the full technical paper: https://lnkd.in/g6n4QjSs 🤗 Explore the model: https://lnkd.in/g6iK9gA5 ⭐ Kudos to all co-authors: Sharath TS, Saurav Muralidharan, Ruisi Cai, Marcin Chochowski, Ameya Mahabaleshwarkar, Yoshi Suhara, Oluwatobi Olabiyi, Daniel Korzekwa, Mostofa Patwary, Mohammad Shoeybi, Jan Kautz, Bryan Catanzaro, Ashwath Aithal, Nima Tajbakhsh, Pavlo Molchanov Stay tuned for what comes next in elastic reasoning models! #AI #LLM #MachineLearning #Efficiency #NVIDIA #EdgeAI #ReasoningModels #Inference #Transformers #SSM #Mamba #Nemotron

Introducing the Nemotron-Personas Collection 💬 A set of multilingual, region-specific synthetic persona datasets created with NVIDIA NeMo Data Designer. Each dataset mirrors real-world demographic and geographic distributions to help you fine-tune and evaluate AI systems without exposing personal data. Now available: 🇺🇸 Nemotron-Personas-USA: 6M personas 🇯🇵 Nemotron-Personas-Japan: 6M personas 🇮🇳 Nemotron-Personas-India: 21M personas All datasets are open source and licensed under CC BY 4.0. Explore on Hugging Face 👉 https://nvda.ws/4ptoAso

NVIDIA Research is making waves at #NeurIPS San Diego 🌊 From workshops and receptions to papers and posters, we’ve got a packed lineup of activities you don’t want to miss 🔗 https://nvda.ws/48rTrj4 Make sure to save the schedule 👇

See NVIDIA DGX Spark in action for local GraphRAG and graph analytics. We will showcase how this demo runs entirely on DGX Spark’s 128GB unified memory, extracting nodes and relationships from local biomedical research papers to build a context-aware Knowledge Graph in ArangoDB.

As #SC25 comes to a close, we're recapping all the highlights. Plus, see the latest on DGX Spark updates, pioneering AI co-scientists, and more 👇

Curious about building your first production-ready AI agent? We’ve got a full hands-on workshop that takes you from the basics to production, so you can get building your own smarter RAG agents ➡️ https://nvda.ws/44hxQHB Here’s some tips to get you started 👇

🎉 Congrats to the winners of NVIDIA DGX Sparks at #SC25! We selected winners during our fireside chat -- each winning a VIP experience and a new DGX Spark. We can't wait to see how they will go #SparkSomethingBig. ✨