NVIDIA Releases Nemotron 3.5 Lightning 30B Open-Weight Model

NVIDIA has released the Nemotron 3.5 Lightning 30B model. This open-weight large language model (LLM) is available under the NVIDIA Open Model License. It is designed for commercial use.

The model features 30 billion parameters. Its architecture is a decoder-only transformer. Nemotron 3.5 Lightning 30B processes up to 128,000 tokens in its context window. It also supports function calling capabilities.

Nemotron 3.5 Lightning 30B was developed by NVIDIA’s Generative AI Products group. The model prioritises on-device and local AI workflows. Its design supports integration into local AI agents and client-side applications. The model is also suitable for edge computing deployments.

The training data consisted of 13.5 trillion tokens. This dataset included both text and code. The training process leveraged NVIDIA GPUs and the NeMo Framework. The model underwent additional fine-tuning. This included reinforcement learning from human feedback (RLHF).

The NeMo framework facilitated various stages. These stages included data curation, distributed training, and model evaluation. The framework also supported inference optimisation for the model. Quantisation to INT4 precision is possible, reducing memory footprint.

NVIDIA provides the model card for Nemotron 3.5 Lightning 30B. This document details model architecture, training methodology, and intended use cases. It also specifies limitations and biases. The license permits commercial deployment and modification. The model represents a continued focus on accessible, performant AI solutions.

Does this open-weight release signal a shift in NVIDIA's model distribution strategy?