Jensen Huang has said variations of the same line for years, but it never sounded this concrete: "We built the computer that runs agents." RTX Spark is the hardware proof. Nemotron 3 Ultra, released June 2, 2026, is the software proof — the most capable open-weights AI model built in the United States, with 550 billion parameters, a 1-million-token context window, and throughput over 300 tokens per second. The benchmark position is only part of the story; the more interesting part is how NVIDIA got here over three years of architectural bets.
From 8 Billion to 550 Billion: The Timeline
The Nemotron family didn't start as a play for frontier models — it started as a tool, the moment NVIDIA realized that dominating AI hardware without a voice in the model ecosystem was a vulnerable position.
- November 2023 — Nemotron-3 8B: an 8-billion-parameter model built for enterprise chatbots, not a frontier play but a declaration of intent.
- February–June 2024 — Nemotron-4: a 15B multilingual model trained on 8 trillion tokens, followed by a 340B family focused on synthetic data generation — this is where NVIDIA learned to build training pipelines at scale.
- January 2025 — Llama Nemotron: reasoning models built on Meta's Llama architecture, the first proof NVIDIA could take a third-party open base and make it competitive with closed frontier models.
- December 2025 — Nemotron 3 Nano: the architectural break, adopting a hybrid Mamba-Transformer design with a latent mixture-of-experts that delivered 4x higher throughput and a 1-million-token context window.
- March 2026 — Nemotron 3 Super: 120 billion total parameters (12 billion active), built to solve the "thinking tax" of running heavy reasoning models for every subtask in a multi-agent system — scored 85.6% on PinchBench, the best open-weights model in its class.
- June 2026 — Nemotron 3 Ultra: 550 billion total parameters (55 billion active per token), available from June 4 on Hugging Face and via NVIDIA NIM, with weights, training data, and replication recipes published openly.
What the Model Does
- Total parameters: roughly 550 billion
- Active parameters per token: roughly 55 billion (mixture-of-experts)
- Architecture: hybrid Mamba-Transformer with latent mixture-of-experts
- Training format: NVFP4, a 4-bit format optimized for Blackwell hardware
- Context window: 1 million tokens
- Throughput: over 300 tokens per second — roughly 5x faster than comparable models
- Cost per agentic task: roughly 30% lower than alternatives
- Intelligence Index (Artificial Analysis): 48 — #1 among US open-weights models
- Availability: Hugging Face and NVIDIA NIM, from June 4, 2026
The throughput and cost numbers come from two decisions: NVFP4 precision, which trains and runs the model at 4-bit specifically tuned for Blackwell hardware so it fits in less memory without meaningful accuracy loss, and a latent mixture-of-experts design that only activates the 55 billion parameters most relevant to a given task instead of all 550 billion. The 1-million-token context window isn't just a benchmark number — it holds roughly 750,000 words, enough for a full codebase or a year of project email, and maintaining that context across hours of autonomous work without losing the thread is a distinct capability from raw reasoning power.
On the Artificial Analysis Intelligence Index, frontier closed models from Anthropic, Google, and OpenAI score around 57. China's Kimi K2.6 from Moonshot AI scores 54. Nemotron 3 Ultra scores 48 — the top US open-weights model, but still 6 points behind the best open-weights model out of China, and the gap to frontier closed models overall has closed from roughly 20 points two years ago to 9 points today. NVIDIA has confirmed a Nemotron 4 is already in development.
The Three-Model Family
Nemotron 3 isn't a single model — it's a three-tier family built for workload routing in agentic systems. Nano (30B total, 3B active) handles high-frequency tasks like classification and summarization, often at the edge. Super (120B total, 12B active) orchestrates multi-agent pipelines and low-latency coordination. Ultra (550B total, 55B active) handles deep reasoning, scientific research, and long-horizon planning. The system evaluates task complexity and routes accordingly, keeping cost low on routine work while reserving maximum capacity for when it genuinely matters.
Why the RTX Spark Synergy Matters
Nemotron 3 and RTX Spark weren't designed by separate teams working in parallel — they were co-optimized from the start. RTX Spark's 128GB of unified memory exists precisely because the 120-billion-parameter Super tier needs that space to run locally without swapping, and the Ultra in NVFP4 format fits in that same memory pool — something no current consumer GPU, including the 32GB RTX 5090, can do. The model was trained to run on this specific hardware, not adapted to it afterward.
An agent running at 300 tokens per second for 8 hours of autonomous work generates roughly 8.6 million tokens — at typical frontier-model cloud pricing, that costs hundreds of dollars a day; running locally, the marginal cost is zero. For small teams, independent researchers, and small businesses, that shift in the economics may matter more than any benchmark score. And because inference happens on-device with no API call and no external log, it satisfies GDPR or CCPA compliance requirements architecturally rather than contractually — while deploying through NVIDIA NIM, vLLM, SGLang, Ollama, or llama.cpp means any developer already working with open-weights models can use it without relearning tools.
The Case For: Intelligence That Belongs to the People Using It
For three years, access to frontier AI came with a hidden cost: every query to a closed model traveled through servers that logged, and potentially used, that content. Nemotron 3 Ultra changes that structurally. With published weights, published training data, and published replication recipes, any developer can audit, modify, fine-tune, and deploy the model inside their own infrastructure — in healthcare, law, defense, and research, that's often the minimum condition for AI to be legally usable at all. The prior Nemotron family had already accumulated over 50 million downloads in the 12 months before this launch, so the community infrastructure to build on it already exists.
The Case Against: Open Weights Aren't Open Source
NVIDIA cultivates an equivalence between "open weights" and "open source" that doesn't hold up. Nemotron 3 Ultra is published under the NVIDIA Open Model License — more permissive than a closed license, but more restrictive than Apache 2.0 or MIT, with redistribution and modification restrictions that don't show up in keynotes. There's also a hardware dependency: the model was trained in NVFP4, optimized for Blackwell, and the benchmark numbers reflect performance on NVIDIA hardware specifically. A free model that only performs at its stated level on one company's hardware is a different kind of lock-in, not the absence of one.
And a 550-billion-parameter open-weights model capable of long-horizon autonomous planning, running locally with no external monitoring layer, is by design a model that lacks the observability a centralized cloud deployment provides. The same property that makes it attractive for legitimate privacy use cases also makes misuse harder to detect — not a reason not to build it, but a reason to be clear-eyed about what "open" actually means here.
What This Means Going Forward
The last four years of AI were defined by centralization — large models in enormous data centers, accessed via API, every interaction passing through a third party. What NVIDIA is building is infrastructure for a decentralized alternative, not because centralized AI is technically inferior, but because there is real, growing demand for intelligence that doesn't depend on connectivity, doesn't carry a variable cost, and doesn't require trusting a third party with your data. Three years from now, the relevant question won't be "which model do I use" — it will be "who controls the model I use."