An AI system that makes its architecture and programming available to everyone. Like regular open-source software, an open-source AI system can be modified for any purpose, and it can be shared with or without modifications. See
open weight AI vs. open-source AI.
Open-source AI comprises the source code, neural network layout (number of layers and neurons), hyperparameters and weights used to program the network along with details about the kinds of data that should be used for training the model (LLM). How the tokenizer splits data into tokens is also provided (see
AI token). However, the actual data to train the model is generally not included. See
AI hyperparameter.
Unlike traditional software systems, the training phase of AI is as important as the execution (inference) side that generates results for the user. See
open source,
TensorFlow,
PyTorch,
Keras,
Rasa,
Amazon SageMaker,
Apache MXNet,
OpenCV,
H2O AI and
OpenAI.