Nvidia is developing the open-source large language model Nemotron 4 as it steps up investment in open-source AI models, The Information reported on Monday, citing sources involved in the work.

The Information said the Nemotron 4 model is expected to have up to more than 1 trillion parameters. It is about twice the size of Nemotron 3 Ultra released in June.

The Information said that even with 1 trillion parameters, it is not smaller in scale than major Chinese open-source models, but Nvidia is strongly emphasizing compression technology that helps smaller models deliver better performance.

Nvidia on Monday also unveiled Nemotron 3.5 Lightning, a small model optimized for running AI agents. It also released free model-routing software. The software helps companies automatically choose the cheapest and most suitable model for each task.

The Information said Nvidia's aggressive push into open-source AI models has also put it in a delicate position in which it must compete with open-source startups it has invested in and major AI companies such as OpenAI, its biggest customer. Nvidia has invested $30 billion in OpenAI.

Kari Briski (커리 브리스키), Nvidia's vice president for generative AI, said in an email, "We invest in Nemotron because we believe every company and country needs accessible frontier open models for safety and innovation."

The Information said the Nemotron consortium includes Reflection, Cursor, Thinking Machines Lab and Mistral, and that Prime Intellect provided 300,000 simulation environments for model training.