A quiet data centre aisle with one server cabinet door standing open, rows of softly lit racks fading into the distance
AI-generated image, not a photograph of a real person or place.
aiopensourcesolutions

WHAT THE DEFINITION SAYS

Weights are not
the whole recipe.

4 FREEDOMS use, study, modify, share

  • Data information
  • Complete code
  • Parameters
  • The three parts of the preferred form

Downloading a model is not the same as open source.

In October 2024 the Open Source Initiative published version 1.0 of its Open Source AI Definition. It asks for more than released weights: enough code and information about the training data to study and rebuild the system.

Updated

Check a release

THE SHORT ANSWER

Open source AI asks
for the full kit.

Released weights let you run a model.
The definition asks what you need to change it.

FREEDOMS

Use, study, modify, share

The definition says an Open Source AI must let anyone use it for any purpose without asking permission, study and inspect it, modify it, and share it, changed or not.

Check a release

WEIGHTS

Weights alone fall short

Weights are the learned parameters of a model. The definition says “Open Source weights” must also come with the data information and code used to derive them.

See the numbers

DATA

Describe the data in full

Not all training data has to be shared, but it has to be described well enough that a skilled person could build a substantially equivalent system.

Why data is described

A QUICK RELEASE CHECK

What did they
actually release?

Pick a typical way an AI model is made available. The bar shows how much of what the Open Source AI Definition asks for is there.

Release check

Choose a release
Online access onlyClosed

You can use the system through someone else's service, but you cannot inspect its components, change it or share it. None of the definition's parts are released.

Weights to download, use limits in the licenceOpen weights, not open source

The parameters are there, but the definition also needs the code and data information. Limits on who may use it, or for what, clash with the freedom to use it for any purpose.

Weights and running code, free to usePartly there

Use and sharing are covered, but studying and rebuilding the system still needs the complete training code and a full description of the training data.

Weights, full code and data information, open termsMatches the three parts

This is the package the definition describes: parameters, the complete code used to train and run the system, and detailed data information, all under OSI-approved terms.

Source: Open Source Initiative, The Open Source AI Definition 1.0 and its FAQ; Open Source Definition, criteria 5 and 6.

A rough guide to the definition's requirements, not a formal assessment of any model and not legal advice on any licence.

TWO NUMBERS WORTH KNOWING

Old rules,
new kind of software.

One number from the 2024 AI definition,
one from the open source rules it builds on.

3 parts

Parts needed to modify an AI system

Data information, the complete code used to train and run the system, and the parameters such as weights. The definition says the preferred form to make modifications must include all three.

PREFERRED FORM · OSAID 1.0

10 criteria

Criteria in the Open Source Definition

The Open Source Definition, version 1.9 of 2007, lists ten criteria for software licences. Two of them say a licence must not discriminate against any person or group, or restrict use in any field of endeavour.

LICENCE TEST · OSI

The OSI announced version 1.0 on 28 October 2024, after what it called a year-long, global, community design process.

FOUR QUESTIONS TO ASK

Before you call
a model open.

Rows of plain archive boxes on grey metal shelving in a calm, well-lit records room
AI-generated image, not a photograph of a real person or place.

These follow the parts the definition names. They help you read a model release page more carefully; they do not replace reading the licence itself.

  1. Can anyone use it for anything?

    Look for limits on who may use the model or for what purpose. The definition asks for use for any purpose without permission.

  2. Is the training code published?

    The definition asks for the complete code used to process the data, train the model and run it, under an OSI-approved licence.

  3. Is the training data described?

    Look for where the data came from, how it was selected and filtered, and where public or third-party data can be obtained.

  4. Are the weights released openly?

    Parameters such as weights must be available under OSI-approved terms, not only for download.

WHY DATA IS DESCRIBED, NOT ALWAYS SHARED

Some data
cannot travel.

A portable external hard drive plugged into a closed laptop on a wooden desk, its small status light glowing
AI-generated image, not a photograph of a real person or place.

The OSI's FAQ explains that it wants Open Source AI to exist where data cannot legally be shared, such as medical AI. It sorts training data into four classes: open, public, obtainable (including for a fee) and unshareable non-public data such as personal information.

Open data must be shared. For the rest, a detailed description is required, so that people can understand the system's biases and others can rebuild a similar model with their own data.

Questions

Questions people ask

Is a model with downloadable weights open source?

Not on its own, under the OSI's definition. Open Source AI also needs the complete training and running code and detailed information about the training data, and the terms must allow use for any purpose.

What are model weights?

The definition describes AI weights as the set of learned parameters that overlay the model architecture to produce an output from a given input.

Does open source AI have to publish all its training data?

No. The OSI's FAQ says open training data must be shared, but other data, such as personal information, can instead be described in detail, including where it came from and how it was selected.

Who wrote the Open Source AI Definition?

The Open Source Initiative, which also maintains the Open Source Definition for software. It announced version 1.0 on 28 October 2024 at the All Things Open conference.

Are there other ways to rate how open a model is?

Yes. A 2024 research paper introduced the Model Openness Framework, a three-tiered ranked system that rates models on the completeness and openness of the code, data and documentation released.

Sources

Where the facts come from

  1. Open Source Initiative: The Open Source AI Definition 1.0
  2. Open Source Initiative: Open Source AI Definition FAQs
  3. Open Source Initiative: announcement of the Open Source AI Definition 1.0 (28 October 2024)
  4. Open Source Initiative: The Open Source Definition
  5. NTIA fact sheet: report on dual-use foundation models with widely available model weights (30 July 2024)
  6. arXiv: The Model Openness Framework (White et al., 2024)

ABOUT THIS NAME

Interested in leasing aiopensourcesolutions.com?

aiopensourcesolutions.com is part of the DoodleSale portfolio. It isn't in an auction right now, but you can ask about leasing it.

Ask about this name