Open-weight

Open-weight models are AI systems that share their pre-trained parameters (the weights) with the public, allowing developers to download, run, and fine-tune them locally.

By distributing the final numerical parameters of a neural network, open-weight models allow developers to bypass the massive compute costs of initial training and run powerful systems locally. While they differ from true open-source software because they rarely disclose the original training data or proprietary codebases, they offer unparalleled control over deployment. Teams can host these models (such as Meta's Llama 3 or Mistral's Mixtral) on private servers, customize them through targeted fine-tuning, and run inference without relying on external APIs.

https://hai.stanford.edu/news/what-open-weight-model