NVIDIA Nemotron 3 Ultra — America's Most Powerful Open-Weights AI Model
NVIDIA's Nemotron 3 Ultra
Arrives as the Most Powerful
Open-Weights AI Model
Built in America
At Computex 2026, Jensen Huang unveiled a 550-billion parameter reasoning giant designed to make agentic AI economically viable — but China's Kimi K2.6 still holds the global lead.
Jensen Huang announced Nemotron 3 Ultra at the Taipei Music Center on June 1, 2026.
On June 1, 2026, NVIDIA CEO Jensen Huang walked onto the stage at the Taipei Music Center — leather jacket firmly in place — and announced what the company is calling its most consequential AI software move to date. The Nemotron 3 Ultra, a 550-billion parameter open-weights language model, is now the crown jewel of NVIDIA's open AI family and, by independent benchmark measurements, the most intelligent open-weights model produced by any American lab.
The announcement was the centerpiece of NVIDIA's Computex 2026 keynote, a presentation that saw Jensen Huang deliberately reframe his company's identity. NVIDIA is no longer simply a chipmaker, he argued. It is a full-stack AI platform company — one that builds the chips, the software infrastructure, and now the models themselves.
Nemotron 3 Ultra is built on a hybrid Mamba-2 Transformer Mixture-of-Experts (MoE) architecture, a design that makes the model radically more efficient than its parameter count suggests. While the model holds 550 billion total parameters, only 55 billion are active for any single token processed — a 90% sparsity rate. NVIDIA calls the underlying mechanism LatentMoE, which compresses tokens into a low-rank latent space before routing them to specialized expert networks. According to the company, this allows the model to call on four times as many expert specialists compared to standard MoE implementations, at the same inference cost.
Paired with multi-token prediction — a technique that generates multiple future tokens in a single forward pass — NVIDIA claims the model achieves over 300 tokens per second. That is a speed figure that places it firmly in the frontier tier for open models. The model also supports a 1 million token context window, enabling it to process entire codebases, lengthy legal documents, or extended research papers in a single inference pass.
One of the most significant aspects of the launch is NVIDIA's decision to release Nemotron 3 Ultra under the NVIDIA Open Model License, which permits commercial use. The weights became available on June 4, 2026 across HuggingFace, OpenRouter, and NVIDIA's own NIM platform. Four checkpoints were released simultaneously: NVFP4 quantized, BF16 instruct, BF16 base, and a generative reward model intended for teams building their own reinforcement learning pipelines. Training data and recipes were also published under the Linux Foundation's OpenMDW-1.1 license.
Over 50 million downloads of the Nemotron 3 family models were recorded in the year leading up to April 2026, a figure that underscores genuine developer appetite for NVIDIA's open model ecosystem. With Ultra now available, that number is expected to accelerate significantly in coming months.
On the Artificial Analysis Intelligence Index, Nemotron 3 Ultra scores 48 — the highest of any US open-weights model. For context, Google's Gemma 4 31B sits at 39, NVIDIA's own Nemotron 3 Super at 36, and OpenAI's gpt-oss-120b at 33. The Ultra's lead over its nearest US competitor is substantial.
However, there is an important caveat that evaluators are careful to highlight. China's Kimi K2.6 scores 54 on the same index — a 6-point gap that analysts describe as meaningful, not marginal. In a global open-weights landscape, the most powerful model remains outside American borders. Whether NVIDIA's next generation — a Nemotron 4 preview was teased during the keynote — can close that gap remains the central competitive question of the second half of 2026.
Running a 550-billion parameter model at production throughput is not a task for consumer hardware. NVIDIA addressed this directly at Computex with two companion hardware announcements. The DGX Station for Windows, powered by the GB300 Grace Blackwell Ultra superchip with 775GB of coherent unified memory, is capable of running Nemotron 3 Ultra without cloud dependency. Partners including HP, Dell, Asus, Supermicro, and MSI are scheduled to bring DGX Station systems to market in Q4 2026.
For developers working with less infrastructure, the newly announced RTX Spark — an Arm-based Grace CPU paired with a Blackwell GPU sharing 128GB of unified LPDDR5X memory — can handle the lighter Nemotron 3 Nano and Nemotron 3 Super (120B) variants. RTX Spark laptops are expected in fall 2026.
NVIDIA's move into open model releases is a calculated strategic bet. By making powerful models freely available, the company ties developer ecosystems to its hardware. Developers who fine-tune, deploy, and build on Nemotron 3 Ultra will naturally gravitate toward NVIDIA's GPU infrastructure when it comes time to scale. The open weights are, in a sense, a long-term sales motion for Blackwell and the generations of chips that follow.
What makes this moment unusual is the breadth of the Computex announcement. Beyond Nemotron 3 Ultra, Jensen Huang also unveiled Cosmos 3 — described as the world's first open Physical AI omnimodel, ranking first across seven robotics benchmarks — and the DGX Station and RTX Spark hardware. The message to the industry was deliberate and unmistakable: NVIDIA intends to compete at every layer of the AI stack simultaneously.
Comments
Post a Comment