NVIDIA has launched Nemotron-4 340 billion, an open family of models developers can use to generate synthetic data for training LLMs for commercial applications across various industries. Nemotron-4 340 billion can be accessed for free and is scalable.

Nemotron-4 340 billion consists of base, instruct, and reward models that form a pipeline. The instruct model generates diverse synthetic data mimicking real-world data, while the reward model filters and grades responses for quality attributes like helpfulness and correctness. The base model can be customized using proprietary data.

NVIDIA claims the open pipeline enables developers to build powerful LLMs by generating high-quality synthetic training data, which is often expensive and difficult to access. The models are optimized for NVIDIA NeMo and TensorRT-LLM for efficient training and inference.