Jensen Huang walked onto the Computex stage in Taipei on June 1, leather jacket on, and announced something that is easy to underrate if you only skim the headline. Nemotron 3 Ultra is not a chip. It is not a GPU. It is a 550-billion-parameter open-weight AI model, and the weights were actually shipped today on Hugging Face, OpenRouter, and NVIDIA’s own NIM service. The company that sells the shovels just released a very good shovel of its own for free and told everyone they could dig wherever they like.
That is the part worth sitting with. NVIDIA makes most of its money selling the hardware that every AI lab on earth needs. It did not have to release a frontier-class open model. The fact that it did says a lot about where NVIDIA thinks the next phase of this goes, and about a quieter strategic shift the company has been making for a while now: NVIDIA is trying to stop being only a chip company and become a full-stack AI platform.
What It Actually Is
Nemotron 3 Ultra is a 550-billion-parameter Mixture-of-Experts model with 55 billion active parameters per token. That distinction matters. A dense 550B model would activate all 550 billion parameters for every single token it generates, which is enormously expensive to run. Mixture-of-Experts routes each token to only the relevant slice of the network, so the model carries the knowledge of a 550B model while paying the compute cost of a 55B one. You get the breadth without the full bill.
The architecture is not a plain Transformer either. It is a hybrid Mamba-Transformer design that combines selective state-space layers with standard attention blocks. The short version of why that matters: pure attention gets expensive fast as context grows, because every token has to attend to every other token. Mamba-style state-space layers handle long sequences far more efficiently. Combining the two lets the model maintain a 1 million-token context window while staying fast, and it scores 94.7 on the RULER long-context benchmark at the full 1 million tokens, while several larger competitors top out at 256K. If you have ever watched a model fall apart once a conversation got long, this is the architecture trying to fix exactly that.
It also relies on NVFP4 quantization, NVIDIA’s 4-bit floating-point format, which helps it achieve roughly 5 to 6 times the inference throughput of comparable open models. On one benchmark configuration, it achieved 5.9x the throughput of GLM-5.1, 4.8x the throughput of Kimi-K2.6, and 1.6x the throughput of Qwen-3.5. It serves over 300 tokens per second, where peer models typically land between 50 and 100. Speed is the whole pitch here, and I will explain why in a second.
This Is a Model Built for Agents, Not Chatbots
Here is the framing that makes Nemotron 3 Ultra make sense. The single-turn chatbot, where you ask a question and get an answer, is yesterday’s use case. The thing everyone is building toward now is the long-running agent: a system that reasons, holds context, calls tools, checks its own results, and grinds through a complex task across many turns without a human babysitting every step.
The problem with long-running agents is that costs add up. Every turn adds tokens, every tool call adds latency, and a task that takes an agent two hundred steps to complete can rack up an alarming inference bill. NVIDIA designed Nemotron 3 Ultra around exactly this economic reality. The insight baked into it is that within any agent workflow, most of the calls are routine and only a critical subset actually demand deep reasoning. Ultra is built to handle those hard calls: sustaining architectural decisions across a long coding session, synthesizing contradictory evidence across hundreds of research sources, or verifying a chip design against thousands of constraints. NVIDIA claims it delivers roughly 30 percent lower cost for complex agentic tasks, which, for a company running agents at scale, is the difference between a feature that ships and one that gets killed in a budget review.
The benchmark numbers back the agentic focus rather than the trivia-quiz focus. It posts 90.0 on PinchBench and 71.9 on SWE-Bench Verified, the latter of which actually measures whether a model can fix real software issues. It hits 56.4 on Terminal Bench 2.1, trailing China’s Kimi-K2.6 at 67.2. And on AA-Omniscience, it records the highest non-hallucination score in its comparison set at 78.7, which is a polite way of saying it is better than most at shutting up when it does not actually know the answer. For an agent that might take an action based on its own output, a model that hallucinates less is worth more than a model that scores two points higher on a knowledge quiz.
The “America’s Smartest Open Model” Caveat
NVIDIA is positioning Nemotron 3 Ultra as the most intelligent US open-weight model available, and on the Artificial Intelligence Index, it lands around 48, which tops the American open-weight field. The jump over its own predecessor is real: Nemotron 3 Super, released back in March 2026 at 120 billion parameters, sits roughly 12 index points lower, and in benchmark terms, that is a big generational leap in a single quarter.
But I want to be honest about the part that NVIDIA’s marketing glosses over: it is the smartest open model built in America, and that qualifier does a lot of work. China’s Kimi-K2.6, from Moonshot AI, still leads the open-weights intelligence ranking outright. So the accurate way to describe Nemotron 3 Ultra is the fastest and most capable US frontier open model, which still trails the best Chinese open model. That is not a knock on NVIDIA’s engineering; it is a snapshot of where the open-weights race actually stands in mid-2026, and the honest version is more interesting than the flag-waving version. The US has been losing the open-weight lead to Chinese labs for over a year, and Nemotron 3 Ultra is NVIDIA’s most serious attempt yet to claw some of it back, backed by a publicly stated five-year, $26 billion plan to fund open-weight AI development.
Actually Open, Which Is the Real Story
What earns my respect here is not the benchmark table; it is the openness. NVIDIA released the weights, the training recipes, and the data pipeline under the NVIDIA Open Model License, which permits commercial use. This is not a weights-only drop with the methodology kept secret. The full pretraining and reinforcement learning recipes are out there, which means other teams can actually learn from how it was built rather than just running the finished artifact.
One technique worth calling out is Multi-Teacher On-Policy Distillation. Instead of training the model against a single larger teacher model, NVIDIA trained Nemotron 3 Ultra with dense feedback from more than ten domain-specific teacher models. The practical upshot is a model that can be continuously improved and specialized for particular domains, rather than a frozen general-purpose blob. For an enterprise that wants to fine-tune the model on its own field, whether that is legal, biotech, or chip verification, that open and adaptable foundation is the actual selling point, more than any single benchmark.
Availability is immediate. As of June 4, the weights are live on Hugging Face, OpenRouter, ModelScope, and NVIDIA NIM. Thirteen early adopters were named at launch, including Accenture, CrowdStrike, Palantir, and Perplexity. The catch, and there is always a catch, is that self-hosting a 550B model requires data center GPU infrastructure. This is not something you run on your gaming PC, and that is, of course, the quiet commercial logic underneath the free model: the more capable open models there are, the more NVIDIA hardware the world needs to run them. The shovel is free. The mine still runs on NVIDIA.
Where It Fits, and What Comes Next
For the local-inference crowd, there is a tidy connection to the rest of NVIDIA’s Computex story. The DGX Spark compact desktop, the same RTX Spark silicon now showing up in the Microsoft Surface Laptop Ultra and the Surface RTX Spark Dev Box, is positioned to run the Nemotron 3 family locally for research teams and power users who want private inference. NVIDIA also confirmed it is already working on Nemotron 4, developed through the Nemotron Coalition, a group of partner organizations, and unveiled Cosmos 3, an open omnimodel aimed at physical AI and robotics. The model release is not a one-off. It is a product line.
So what do I actually make of it? Nemotron 3 Ultra is a strong, fast, openly licensed model that is honest about being optimized for agents rather than for topping a single intelligence leaderboard. It does not beat China’s best open model, and NVIDIA’s framing works hard to avoid saying that out loud. But the openness is real, the agentic engineering is thoughtful, and the strategic message is louder than the model itself: the company that owns the AI hardware layer now wants to own the open model layer too. Whether that is generous or just very long-term self-interest depends on how cynical you are feeling. It is probably both, and the free frontier-class weights are useful either way.
Sources: NVIDIA Technical Blog, NVIDIA Newsroom, Artificial Analysis, MarkTechPost, Decrypt, ChatForest, explainx.ai, The News. Reported from Computex 2026, Taipei, June 4, 2026.