Meta’s Llama 3.1 illustrates the appeal of privately operated AI. Meta says the model weights are available to download and that developers can customize the models, train them on new datasets, conduct additional fine-tuning, and run them on premises without sharing data with Meta. For the 405B model, Meta also describes work to support production inference at that scale.
Those capabilities give an organization more control over where its data and workloads go. They do not, by themselves, answer a different question: what terms accompany the model being used?
NVIDIA’s Llama-3.1-Minitron-4B-Width-Base makes the distinction concrete. NVIDIA’s model card says the company created Minitron by pruning Meta’s Llama 3.1 8B and then continuing training with distillation using 94 billion tokens. The card releases the resulting model under the NVIDIA Open Model License Agreement.
The same repository also preserves a NOTICE stating that Llama 3.1 is licensed under the Llama 3.1 Community License and identifying Meta as the copyright holder. A team examining Minitron therefore encounters terms declared for the downstream model alongside a notice connected to the upstream model from which it was made.
The presence of those two agreements does not establish that they conflict, and it does not determine which provisions apply to a particular use. It shows why reading only the headline license can leave the factual picture incomplete. The model’s lineage and the notices retained in its repository are also relevant to understanding the asset.
For this specific Minitron model, the FastCatalog record brings that information together. It identifies the NVIDIA agreement for the downstream component, shows Meta’s Llama 3.1 8B as a direct upstream model, and associates the Llama 3.1 Community License with that upstream component. The record does not decide which terms govern a proposed use; it makes the underlying fact-finding easier.
Private infrastructure can clarify where AI work happens. The Minitron example shows why understanding what is being run may require looking beyond the downloaded files to the asset’s upstream history.