What does the Open Source AI Definition require?

Version 1.0 of the Open Source AI Definition says an Open Source AI system is available under terms that let people use it for any purpose, study how it works, modify it, and share it, with or without changes.

The precondition is access to the preferred form for making modifications.

For a machine-learning system, that preferred form has three parts.

Data information must be detailed enough that a skilled person can build a substantially equivalent system, including the provenance, scope, selection, labeling, and filtering of the training data, plus where to obtain the public and third-party data.

Code must include how the data was processed and how training and inference are run. Parameters are the weights and related settings, which may include intermediate checkpoints.

OSI-approved terms apply to the data information, the code, and the parameters.

How is that different from open weights?

The same page says an AI model is the architecture, the parameters, and the inference code, while weights are the learned parameters.

It then says “open source models” and “open source weights” must still include the data information and the code used to derive those parameters.

A release that posts a weight file and keeps the training recipe private can be an open-weight release under the publisher’s own wording. It is not, on this definition, open source AI.

The license checklist is the place to record the exact file you reviewed. This article is not a legal opinion on a particular license.

What you can have without meeting the OSI definition
What you receivedWhat it tells youWhat it does not tell you
Weight fileYou can run that checkpoint if the terms allow itThat you may retrain or redistribute it
Inference codeYou can see how a forward pass is runThat the training code and data description are included
A marketing label of “open source”The publisher’s own descriptionThat the release meets the OSI definition

What should you check before you deploy?

Write down the model, the version, the license file, and whether the training-data description and training code are in the release. Recheck when you upgrade.

A weight file that was free to download last month can ship under a different grant in the next checkpoint.

Common questions

Is an open-weight model the same as an open-source LLM?

Not under the OSI definition. Open weights are the parameters.

Open source, in that definition, also includes the data information and the code used to produce them.

Does Apache 2.0 on a weight file meet the definition by itself?

No. The license on the file you downloaded does not supply missing training data information or training code.

Sources & further reading

  1. Open Source Initiative: Open Source AI Definition 1.0 ↗
How this story was made

Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.

Our editorial standards
Back to all stories