NVIDIA Launches Nemotron 3 Open Models at GDC — 120B Parameters With 5x the Throughput
NVIDIA's Nemotron 3 family ships in Nano, Super, and Ultra sizes with up to 1M-token context, already adopted by CrowdStrike, Cursor, Perplexity, and Zoom.
Open Models, Enterprise Scale
NVIDIA used its GDC 2026 keynote for more than just graphics — it also unveiled the Nemotron 3 family of open-weight AI models, representing the company's most aggressive push into the foundation model space. The family ships in three sizes: Nano (for edge and mobile), Super (the workhorse), and Ultra (the frontier competitor), all available under permissive commercial licenses.
Nemotron 3 Super is the standout. It's a 120-billion-parameter model that uses a sparse mixture-of-experts architecture with only 12 billion parameters active per inference pass. The result is frontier-class reasoning at a fraction of the compute cost — NVIDIA claims up to 5x higher throughput and 2x higher accuracy compared to the previous Nemotron Super generation. The model supports a 1-million-token context window natively, putting it in the same league as Gemini 3.1 Pro and GPT-5.4 for long-document processing.
Who's Already Building With It
The adoption list reads like a who's who of AI-powered products. CrowdStrike is integrating Nemotron 3 into its threat detection pipeline. Cursor — the AI code editor — is using it as an alternative reasoning backend. Perplexity is deploying it for search synthesis. ServiceNow is embedding it in enterprise workflow automation. And Zoom is using Nemotron 3 Nano for on-device meeting summarization that runs entirely locally on laptops.
That last use case highlights the real strategic play: NVIDIA isn't just competing in the cloud model market. By shipping Nano and Super variants optimized for its own GPU hardware, NVIDIA is building a model ecosystem that makes its chips the natural deployment target from edge devices to data centers.
What This Means for the Open Model Landscape
Nemotron 3 arrives at an interesting moment for open-weight models. Meta's Llama 3.3, Mistral's Large 2, and DeepSeek's V3 have all pushed the boundaries of what's possible outside proprietary model providers. NVIDIA's entry adds a new dimension: hardware-optimized open models from the company that makes the GPUs everyone trains on.
The combination of open weights, commercial licensing, hardware optimization, and enterprise-grade support makes Nemotron 3 a compelling option for organizations that want frontier capabilities without vendor lock-in to a single model API provider. Whether it displaces Llama or Claude in production deployments remains to be seen — but the competition is exactly what the ecosystem needs.
Sources: NVIDIA Newsroom (March 10, 2026), VentureBeat (March 10, 2026), The Verge (March 10, 2026)