NVIDIA introduced Nemotron 3 Super — an open model with 120 billion parameters

NVIDIA Unveiled Nemotron 3 Super — An Open Model with 120 Billion Parameters

The gap between closed commercial systems and open language models is narrowing at a pace that would have seemed impossible just a few years ago. NVIDIA, a company whose name has long become synonymous with hardware superiority in the era of artificial intelligence, is taking increasingly confident steps into the software space as well. Its new release — Nemotron 3 Super with 120 billion parameters — claims not just a place in the leadership table, but a redefinition of the very logic behind building open models for agentic tasks.

To understand the significance of this release, one must look at the context. Over the past two years, open models have developed rapidly, yet they have always lagged behind closed counterparts in several key characteristics: performance on complex multi-step tasks, inference speed, and the ability to work efficiently in multi-agent scenarios. GPT-4, Claude, and Gemini maintained a quality lead for a long time precisely in these areas. Nemotron 3 Super was created as a response to this gap — not a compromise between openness and quality, but an attempt to eliminate the very necessity of such a compromise.

The technical solution underlying the model deserves special attention. NVIDIA applied a hybrid architecture in which the Mamba mechanism — an efficient alternative to the classical Transformer when working with long sequences — is combined with the traditional Attention mechanism. Layered on top of this is the Mixture-of-Experts approach, which allows activating only part of the parameters with each call to the model. The result is a fivefold increase in throughput compared to architectural analogues in the same weight class. This is not a marketing metric: in real agent systems, where the model processes dozens of parallel requests and manages tool chains, inference speed often becomes the bottleneck of the entire system.

In NVIDIA's product lineup, Nemotron 3 Super occupies a well-considered intermediate position. The compact version with 30 billion parameters is aimed at deployment with limited resources and local applications. Larger solutions are for tasks requiring maximum reasoning depth. Nemotron 3 Super fills the space between them: powerful enough for complex corporate scenarios, efficient enough to not require massive infrastructure expenditures. This positioning suggests that NVIDIA is designing not a separate product, but an ecosystem in which each component solves its own task without redundancy.

The consequences of this release extend far beyond technical discussion. First, it increases pressure on OpenAI, Anthropic, and Google in the corporate client segment. Companies that previously chose closed APIs for quality reasons now have an argument for switching to open solutions: full data control, the ability to fine-tune, and the absence of dependence on an external provider. For privacy-sensitive industries — finance, healthcare, government sector — this could be a decisive factor in choosing infrastructure.

Equally important is the signal that Nemotron 3 Super sends to the research community. Open weights mean the possibility of reproduction, audit, and improvement — something that closed models are fundamentally deprived of. If the Mamba-Attention hybrid architecture indeed delivers the claimed efficiency gains, it could become a new standard for the next generation of open systems. NVIDIA is essentially publishing not just a model, but an architectural reference point.

The release of Nemotron 3 Super marks a moment when open AI ceases to be synonymous with "good enough" and begins to claim the role of best-in-class. NVIDIA skillfully converts its dominance in the hardware space into software authority, forming a complete stack — from chips to model weights. For the market, this means intensified competition and, as a result, accelerated progress. For developers — expanded real choice. And for the entire industry — confirmation that the most exciting race in artificial intelligence today is unfolding not behind closed laboratory doors, but in open access.