NVIDIA unveils Nemotron 3 Super, an open 120B-parameter model

NVIDIA unveils Nemotron 3 Super, an open 120B-parameter model

The gap between closed commercial systems and open language models is closing at a pace that would have seemed impossible just a few years ago. NVIDIA, whose name has long been synonymous with hardware dominance in the AI era, is now making increasingly confident moves in software as well. Its new release, the 120-billion-parameter Nemotron 3 Super, aims for more than a spot on the leaderboard: it sets out to redefine how open models for agentic tasks are built.

To understand why this release matters, consider the context. Over the past two years, open models have advanced rapidly, yet they have consistently trailed their closed counterparts on several key metrics: performance on complex multi-step tasks, inference speed, and the ability to work effectively in multi-agent scenarios. GPT-4, Claude, and Gemini held a clear quality lead in exactly these areas for a long time. NVIDIA built Nemotron 3 Super as a response to that gap. Rather than a compromise between openness and quality, it is an attempt to eliminate the need for that compromise altogether.

The technical design behind the model deserves a closer look. NVIDIA uses a hybrid architecture that combines Mamba, an efficient alternative to the classic transformer for long sequences, with a traditional Attention mechanism. On top of that sits a Mixture-of-Experts approach, which activates only a fraction of the parameters for each request. The result is a fivefold throughput gain over architectural peers in the same weight class. This is not a marketing number: in real agentic systems, where a model handles dozens of parallel requests and orchestrates tool chains, inference speed is often the bottleneck for the entire system.

Nemotron 3 Super occupies a deliberate middle position in NVIDIA's product lineup. The compact 30-billion-parameter version targets resource-constrained deployments and local applications. Larger models handle tasks that demand maximum reasoning depth. Nemotron 3 Super fills the space between them: powerful enough for complex enterprise scenarios, yet efficient enough to avoid massive infrastructure costs. This positioning suggests NVIDIA is designing not a standalone product but an ecosystem in which each component does its own job without redundancy.

The implications of this release reach well beyond technical debate. First, it puts more pressure on OpenAI, Anthropic, and Google in the enterprise segment. Companies that previously chose closed APIs for quality reasons now have a case for switching to open alternatives: full control over their data, the ability to fine-tune, and no dependence on an external provider. For privacy-sensitive industries such as finance, healthcare, and government, this could prove decisive when choosing infrastructure.

Just as important is the signal Nemotron 3 Super sends to the research community. Open weights make reproduction, auditing, and further development possible, which closed models fundamentally cannot offer. If the hybrid Mamba-Attention architecture really delivers the claimed efficiency gains, it could become the new standard for the next generation of open systems. In effect, NVIDIA is publishing not just a model but an architectural reference point.

The Nemotron 3 Super release marks the moment when open AI stops meaning "good enough" and starts competing for best in class. NVIDIA is skillfully converting its hardware dominance into software authority, building a full stack from chips to model weights. For the market, this means fiercer competition and, as a result, faster progress. For developers, it means more real choice. And for the industry as a whole, it confirms that the most exciting race in AI today is playing out not behind the closed doors of labs, but in the open.