MMOOZZGG.com

Glossary

What is open source AI?

In short

Open source AI is an approach in which a model is published so that anyone can freely use, study, change and share it: with the weights, the code and information about the training data.

Language models · open source AI

MMOOZZGG.com editorsChecked: 1 min read

For a child

Imagine a cook who not only gives you a piece of pie but also the recipe and the list of ingredients, and lets you bake the same pie at home and change the filling. That is generous: anyone can learn and do it their own way. Smart programs can be like that too. Some authors show everything about how they made their program, while others only let you try it.

In plain words

You need three things to build a neural network: code, training data and the final weights, which are the learned numbers. A project is open in the full sense when everything needed to repeat and change the model is available, and the license lets anyone do this for any purpose. Models are often called open when only their weights are published, while the data and training code stay closed or the license limits use. This is useful, but it is not the same thing.

For business

For a company, openness means independence: you can run the model on your own servers, check what it was trained on and fine-tune it without the vendor’s permission. This helps with data storage requirements in Russia and lowers the risk that a service shuts down or changes its prices. But an open system is not free to run: you need hardware, engineers, updates and security checks. Before choosing, a lawyer reads the license: the word “open” in a description does not always allow commercial use. Decide separately who will update the model and close vulnerabilities. Articles on sovereign AI help compare the cost of running a model yourself with the cloud.

For a specialist

Open source AI, as defined by the Open Source Initiative (Open Source AI Definition 1.0), is a system made available on terms that grant four freedoms: to use it for any purpose without asking permission, to study how it works, to modify it and to share it. This requires the preferred form for making changes: data information sufficient for a skilled person to build a substantially equivalent system; the complete source code for training and inference under OSI-approved licenses; and the parameters, to which checkpoints from key training stages can be attached. The Model Openness Framework (White and co-authors, 2024) proposes a three-tier classification of openness by code, data and documentation and criticizes restrictive model licenses that are called open source. An example of a fully open project is OLMo (Groeneveld and co-authors, 2024), which published its training data, training code and evaluation code. Models with open weights only do not formally belong to this category.

Example

A clinic chooses a model to process medical records. It takes a project with open data and training code to check whether the training set held other people’s personal data, and runs the model inside its own network.

Related terms

Questions and answers

How is open source different from open weights? With open weights you get a ready model, but you do not know exactly how it was trained, and you are bound by the license terms. A fully open project also gives you the training code, information about the data and the right to any use.

So open weights are only part of openness.

Can you use open models for commercial purposes? If the license meets the definition of open source AI, yes, with no limits on purpose. But many models called open have licenses that ban certain fields or set conditions for large services.

So read each license before rollout.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us