NVIDIA is strengthening its open Large-Scale Language Model (LLM) strategy, placing 'efficiency' at the forefront as a core value of the artificial intelligence (AI) era.

Bryan Catanzaro, Vice President of Applied Research at NVIDIA, emphasized during his keynote speech at 'NVIDIA Nemotron Developer Days Seoul 2026' held at D.CAMP in Mapo-gu, Seoul on the 21st, that “efficiency is intelligence, and the key is to implement higher levels of artificial intelligence with limited computing resources.”

Vice President Catanzaro described the stages of AI development as an evolutionary process leading from conversational models to reasoning models, and finally to agents.

The explanation is that the agent is a system that goes beyond a single model to combine memory, multimodal input, access to file and messaging tools, and computer skills, dramatically expanding an individual's productivity and problem-solving capabilities.

He described it as “an era where every individual becomes a CEO running their own research institute.”

At the center of these changes is NVIDIA's open LLM family, 'Nemotron'There is 'Nemotron'.

Vice President Catanzaro assessed that AI no longer operates as a single model but is evolving into a structure where multiple models and systems are combined.

Accordingly, they explained that four 'scaling laws'—pre-training, post-training, the inference phase, and agent operation—are operating simultaneously, explosively driving up computing demand.

He said, “Computing is intelligence, and the higher the efficiency, the more intelligence can be gained.”

In accordance with this philosophy, Nemotron adopted an open strategy that broadly discloses not only the model itself but also the training dataset, training techniques, hyperparameters, and software.

This is explained as a choice intended to enable developers and companies to specialize AI to suit their respective needs while simultaneously maintaining data and technology sovereignty.

Nemotron also has a direct impact on NVIDIA's hardware design.

To efficiently process large-scale expert mix (MoE) models, an ultra-low latency, high-bandwidth interconnect was designed, and in the latest GPU generation, 4-bit numeric representation (NVFP4) was introduced to simultaneously improve energy efficiency and performance.

In fact, it was introduced that the Nemotron-3 series models were pre-trained using only 4-bit operations.

At the event, the progress of the third generation of Nemotron was also revealed.

Following the 30 billion parameter Nemotron Nano and the 120 billion parameter Nemotron Super, the super-large model 'Ultra' is in the post-training stage, and a multimodal inference model encompassing images, voice, video, and text is also scheduled to be released soon.

Vice President Catanzaro“Nemotron Super has demonstrated strengths over competing models in terms of long-term context reasoning and speed,” he said.

The release of a dataset targeting Korea also drew attention.

NVIDIA has released a dataset of 7 million synthetic personas reflecting Korea's demographic, linguistic, and cultural statistics.

It is 'privacy-designed' data that reflects actual distributions without including personal information, intended to support domestic developers in creating AI specialized for the Korean environment.

Vice President Catanzaro stated, “NVIDIA’s goal is not to control AI, but to accelerate the ecosystem,” while also emphasizing the company’s commitment to collaborating with domestic companies and research institutions.

He said, “Korea’s AI momentum is very impressive, and I look forward to building an open and accelerated future of AI together.”


Meanwhile, at 'Nemotron Developer Days Seoul 2026' hosted by NVIDIA, events such as building your own AI agent with OpenClaw & NemoClaw, live demos with NVIDIA experts, and DGX Spark purchase consultations and on-site purchases were held.

The NVIDIA Nemotron technical team engaged with participants through practical technical sessions covering model development, data construction, and AI agent design.