NVIDIA Unveils Nemotron 3 Ultra, Targets the Next Wave of AI Agents
Jensen Huang used his Taipei keynote to position NVIDIA as a full-stack AI platform, installing Nemotron 3 Ultra atop its open model family with a latent MoE trained in NVFP4 and a claimed 5x throughput uplift.
By Ryan Merket ·
Why it matters
NVIDIA hinting at a new Nemotron release will draw developer attention, but real adoption hinges on concrete details like license, cost, and performance on NVIDIA GPUs.
NVIDIA has unveiled Nemotron 3 Ultra, a new flagship open model aimed squarely at the emerging market for AI agents, autonomous workflows, and enterprise reasoning systems.
The announcement came during a keynote from NVIDIA CEO Jensen Huang, who presented benchmark results positioning Nemotron 3 Ultra among the strongest open models currently available.
https://x.com/NVIDIAAI/status/2061305524700758050
Rather than focusing solely on traditional AI benchmarks like math competitions or coding exams, NVIDIA highlighted metrics tied directly to real-world agent performance, signaling where the company believes the industry is headed next.
NVIDIA's bet: AI is moving beyond chat
The presentation framed Nemotron 3 Ultra as a model built for what NVIDIA calls "Frontier Smart" applications.
According to benchmark data shown on stage, the 550-billion-parameter model posted competitive results across a range of agent-focused evaluations.
| Benchmark | Nemotron 3 Ultra | GLM 5.1 | Kimi K2.6 | Qwen 3.5 |
|---|---|---|---|---|
| Agent Productivity | 91% | 84% | 91% | 89% |
| Long-Horizon Planning | 33% | 40% | 29% | 30% |
| Coding | 54% | 64% | 67% | 53% |
| Instruction Following | 82% | 77% | 74% | 78% |
| Knowledge Work | 1,448 | 1,594 | 1,508 | 1,192 |
| Professional Work Tasks | 56% | 46% | 56% | 53% |
| Long Context | 95% | N/A | N/A | 90% |
While Nemotron did not lead every category, NVIDIA emphasized its balanced performance across agent productivity, instruction following, enterprise tasks, and long-context reasoning.
Those are increasingly the capabilities developers care about as AI systems evolve from question-answering tools into software that can complete complex workflows on behalf of users.
A benchmark suite designed for agents
One of the most notable aspects of the launch wasn't necessarily the scores themselves.
It was the benchmarks NVIDIA chose to showcase.
For much of the past three years, AI companies have competed primarily on coding tests, mathematical reasoning, and standardized evaluation suites.
Nemotron's launch presentation looked different.
Metrics such as:
- Agent Productivity
- Long-Horizon Planning
- Professional Work Tasks
- Long Context
- Instruction Following
suggest NVIDIA is optimizing for systems that can act, execute, and collaborate rather than simply respond.
That aligns with a broader shift occurring across the AI industry.
Founders, investors, and enterprise buyers are increasingly focused on agentic systems capable of handling customer support, software development, research, operations, and other multi-step workflows that previously required human intervention.
NVIDIA is climbing the AI stack
For years, NVIDIA occupied a relatively simple position in the AI ecosystem.
The company sold the GPUs powering everybody else's models.
That role has become one of the most profitable businesses in technology history, but Nemotron shows NVIDIA's ambitions extend far beyond hardware.
The company is increasingly building:
- Foundation models
- Agent frameworks
- AI infrastructure
- Enterprise AI tooling
- Inference platforms
- Developer ecosystems
The logos displayed alongside the Nemotron launch reinforce that strategy.
Partners and ecosystem participants shown during the presentation included Cursor, LangChain, Nous Research, Perplexity, Black Forest Labs, Reflection, NAVER, Mistral, and others.
Many of these companies sit at the center of the emerging agent economy.
Rather than serving as a supplier behind the scenes, NVIDIA appears increasingly interested in becoming a platform company that powers both the hardware and software layers of AI.
Entering a crowded battlefield
Nemotron arrives at a moment when open-weight AI competition is becoming increasingly intense.
Over the last year, companies including DeepSeek, Moonshot AI, Alibaba, and Zhipu have released increasingly capable reasoning models that have challenged assumptions about where frontier AI innovation originates.
Chinese labs in particular have compressed the performance gap while dramatically lowering costs, forcing U.S. companies to rethink their competitive positioning.
The benchmark comparisons shown by NVIDIA directly acknowledge that reality.
Rather than comparing Nemotron exclusively against Western models, the company placed its system alongside GLM 5.1, Kimi K2.6, and Qwen 3.5, all of which have emerged as major players in the global open-model ecosystem.
That comparison reflects a growing consensus inside the industry: the competition for frontier AI is now truly global.
Why it matters
The significance of Nemotron 3 Ultra extends beyond another model launch.
The benchmark categories highlighted during the announcement provide a glimpse into how NVIDIA sees the future of AI.
The next generation of AI systems may not be judged primarily by how well they answer questions.
Instead, they may be evaluated by how effectively they can:
- Plan over long time horizons
- Maintain context across large workflows
- Coordinate multiple actions
- Execute professional tasks
- Operate as autonomous agents
If that vision proves correct, Nemotron 3 Ultra is less a chatbot competitor and more a foundation for digital workers.
For NVIDIA, the launch marks another step in a broader transformation from GPU manufacturer to full-stack AI platform.
And if the company's performance claims hold up under independent evaluation, Nemotron 3 Ultra could become one of the most important open models available to developers building the next generation of AI agents.