Nvidia Emerges as a Leading Model Creator with Nemotron 3

Nvidia has generated significant revenue by providing chips to businesses focused on artificial intelligence. However, the chip manufacturer has now taken steps to position itself as a serious model creator by launching a series of state-of-the-art open models, accompanied by data and tools designed to assist engineers in utilizing them.

This initiative comes at a time when AI companies such as OpenAI, Google, and Anthropic are creating more advanced chips independently, potentially serving as a safeguard against these firms shifting away from Nvidia’s technology in the future.

Open models play a vital role in the AI landscape, with numerous researchers and startups leveraging them for experimentation, prototyping, and development. While OpenAI and Google offer some smaller open models, they update them less frequently compared to their competitors in China. Consequently, open models from Chinese firms are currently much more in demand, as indicated by data from Hugging Face, a platform for hosting open-source projects.

According to benchmark scores released prior to availability, Nvidia’s new Nemotron 3 models rank among the finest that can be downloaded, customized, and operated on personal hardware.

“Open innovation is the cornerstone of AI advancement,” stated CEO Jensen Huang, addressing the news. “With Nemotron, we’re evolving advanced AI into an open platform that provides developers the transparency and efficiency necessary for building scalable agentic systems.”

Nvidia is adopting a more transparent strategy compared to many of its American competitors by releasing the data utilized to train Nemotron, facilitating easier model modifications for engineers. The company is also providing tools for customization and fine-tuning, which includes a new hybrid latent mixture-of-experts architecture touted by Nvidia as particularly suitable for developing AI agents capable of performing tasks on computers or the internet. Furthermore, the company will introduce libraries that enable users to train agents through reinforcement learning, which involves simulating rewards and penalties for the models.

The Nemotron 3 models are available in three configurations: Nano, with 30 billion parameters; Super, featuring 100 billion; and Ultra, which boasts 500 billion. A model’s parameters roughly reflect its capability as well as the complexity involved in running it. The largest models are so demanding that they require multiple racks of expensive hardware.

Model Foundations

Kari Ann Briski, vice president of generative AI software for enterprise at Nvidia, noted that open models are crucial for AI developers for three reasons: they increasingly require customization for specific tasks; they benefit from delegating queries to various models; and they yield more intelligent responses after being trained to perform a form of simulated reasoning. “We believe open source is fundamental to AI innovation, continuously driving the global economy forward,” Briski remarked.

Meta, the social media giant, introduced the first advanced open models under the label Llama in February 2023. However, as competition has heightened, Meta has indicated that its upcoming releases may not be open source.

This development is part of a broader trend in the AI sector. Over the past year, US companies have veered away from transparency, becoming more secretive about their research and hesitant to reveal their latest engineering developments to competitors.