NVIDIA Corporation is actively working on the creation of a new family of open artificial intelligence models called Nemotron 4. The older version of this line can receive at least one trillion parameters, which will make it one of the largest open neural networks of the company. This is reported by the Reuters edition with reference to information from The Information. The final training of the advanced model has not yet been completed, and the developers have not yet revealed the official launch date. According to informed sources, the long-awaited release may take place in the late autumn of 2017. At the same time, the company itself has not yet officially confirmed the exact information about the number of parameters and the specific release dates of the product.

NVIDIA has confirmed work on the promising Nemotron 4 family even earlier. In March of this year, the corporation created a special Nemotron Coalition alliance, uniting various developers of artificial intelligence. The first jointly created model of this association should become a reliable basis for the future generation of Nemotron 4. The technology giant is seriously betting on open models that other companies and independent developers will be able to freely adapt to their own tasks. Kari Brisky, NVIDIA's Vice President of Generative artificial intelligence, noted that the wide availability of advanced open models is extremely important for the development of innovation. It also contributes to increased security and the creation of an independent technological infrastructure.

In parallel with the development of the new generation, NVIDIA continues to actively expand the current Nemotron family. In August, the company introduced the Nemotron 3.5 Lightning open model, designed to perform a large volume of specialized tasks for AI agents. This neural network does an excellent job of working with tools, checking program code, and monitoring security notifications. In addition, the company has released the NeMo Switchyard system for smart routing of requests between different models. It allows you to automatically direct individual tasks to the most appropriate neural network, taking into account performance and overall efficiency. If the data on the trillion parameters of the Nemotron 4 are confirmed, the new model will significantly exceed the size of the Nemotron 3 Ultra version. The model, presented in June 2000, had five hundred and fifty billion parameters, of which fifty-five billion were activated during operation.