Open-Source AI
Open-source AI is AI software released under terms that let anyone use, study, modify and share it, and the phrase has become contested because a trained model is more than source code. A model is a bundle of code, a training recipe, training data and the learned numbers called weights, and each piece can be released or withheld independently. In October 2024 the Open Source Initiative published the Open Source AI Definition 1.0 to say what must be shared before an AI system can honestly be called open source, after about two years of consultation.
The OSI definition carries the four classic freedoms of free software over to AI systems: to use for any purpose, study how it works, modify it, and share it with or without modifications. To make those freedoms real, it requires access to the preferred form for making modifications, which it breaks into three parts: the model parameters; the complete source code used to train and run the system; and sufficiently detailed information about the training data so that a skilled person could build a substantially equivalent system. Notably it does not demand the raw training data itself, because some data cannot be legally redistributed, such as copyrighted or private material. This compromise was criticized from both sides: some argued that withholding the data makes studying a model impossible, while others worried that stricter rules would exclude almost every useful model.
Open weights is a narrower and more common thing than open source. An open-weight release publishes the trained parameters so that anyone can download, run, fine-tune and often redistribute the model, but it may omit the training data, the data-processing code and the training logs, and it often comes with a license that has conditions. Meta's Llama family is the best-known case. Llama 2, released in July 2023 with Microsoft as a launch partner, used a community license that required companies with more than 700 million monthly active users to ask Meta for permission, attached an acceptable use policy, and initially barred using the outputs to improve other language models. The OSI has stated that Llama licenses do not meet its definition, because restrictions on users and fields of use conflict with the freedom to use for any purpose, even though Meta markets the models as open. By contrast, models such as Mistral 7B were released under the permissive Apache 2.0 license, and DeepSeek released its R1 model under the MIT license in 2025, though both still publish weights without the full training data.
OLMo shows what a maximally transparent release looks like. The Allen Institute for AI released OLMo in February 2024 as a fully open language model, publishing the weights alongside the Dolma pretraining corpus of about three trillion tokens, the code for data preparation, training and evaluation, intermediate checkpoints and training logs, all under the Apache 2.0 license. Such releases let researchers reproduce results, audit the effect of data choices and study how capabilities emerge during training, none of which is possible with weights alone. Other projects with substantial openness include EleutherAI's Pythia suite and the BigScience BLOOM model, which used a Responsible AI License that restricts certain uses and is therefore not open source in the OSI sense.
Why the label matters is a mix of law, security and marketing. Companies gain goodwill and ecosystem adoption by using the term open, and some European Union rules in the AI Act give limited exemptions to free and open-source models, so the definition has regulatory consequences. Supporters of openness argue that public weights allow independent safety research, local and private deployment, and less concentration of power; skeptics point out that released weights cannot be recalled and that safety fine-tuning can be removed by downstream users. A frequent pitfall is equating free to download with free to use for anything, when the license text is what actually grants or withholds rights. Another is assuming that open-source code for an inference runtime makes the model open, when the weights carry their own terms.
Read open-source AI as a spectrum with a formal threshold: downloadable weights give you something to run, but the OSI definition asks for the freedom to study and rebuild, which means code, parameters and meaningful information about the data, and most models marketed as open today sit somewhere short of that line.