Nvidia launches Nemotron 3 Ultra at Computex 2026

Nvidia Launches Nemotron 3 Ultra 500B AI Model at Computex 2026

Nvidia CEO Jensen Huang unveiled the Nemotron 3 Ultra, a 550-billion-parameter open-weights AI model, during his keynote at Computex 2026 in Taipei on Monday, marking the company’s most aggressive move yet into enterprise AI software.thenews

A New Flagship for Agentic AI

The model uses a mixture-of-experts architecture with roughly 55 billion parameters active per token and 90% sparsity, making it far more efficient than its total parameter count suggests. According to Artificial Analysis, it scores 48 on the firm’s Intelligence Index, placing it ahead of all other U.S. open-weights models including Google’s Gemma 4 31B at 39, though still trailing China’s Kimi K2.6 at 54.artificialanalysis

Nvidia said Nemotron 3 Ultra delivers more than 300 output tokens per second — three to six times faster than peer models of similar intelligence such as DeepSeek and Moonshot, which typically serve 50 to 100 tokens per second. The company claims it reduces costs by roughly 30% for complex agentic tasks compared with leading alternatives.coinpedia

The model is tuned specifically for multi-step reasoning, planning, and self-correction over long task horizons — capabilities central to autonomous AI agents that can execute workflows without human intervention.beam

A Full-Stack Enterprise Play

Nemotron 3 Ultra sits atop a new three-tier Nemotron 3 family that includes the mid-range Super and the lightweight Nano Omni, a multimodal edge model that unifies vision, audio, and language for on-device agents. Alongside the models, Huang introduced NemoClaw, an orchestration framework for agent planning and delegation, and OpenShell, a security and governance runtime layer.thenews

The announcements came as part of a broader keynote that also featured the Vera CPU — a processor purpose-built for agentic AI workloads that Nvidia claims delivers twice the efficiency of traditional x86 server chips — and the RTX Spark, a consumer superchip combining an Arm CPU and Blackwell GPU with up to 128GB of unified memory.beam

Positioning Against Rivals

The suite of announcements positions Nvidia not merely as a chipmaker but as a full-stack AI platform company competing with the likes of OpenAI, Google, and Meta in model development while retaining its dominance in hardware. The Vera Rubin platform, first detailed at GTC in March, is now attracting early cloud adopters including AWS, Google Cloud, and Microsoft.nvidia

Computex 2026, running June 1–5 under the theme “AI Together,” serves as the stage for Nvidia’s vision of a world where AI agents run everywhere from data centers to laptops — with the company supplying every layer of the stack.nvidia