What Open Source AI Means: An Artifact Test

What Open Source Means for an AI Release

Open source AI should mean more than a public paper, visible code, or downloadable weights. A useful review asks whether people can use, study, modify, and share the system for any purpose, then checks whether the code, parameters, and information needed to make modifications are actually available under suitable terms.

That standard comes from the Open Source AI Definition 1.0, released by the Open Source Initiative in October 2024. It gives writers and builders a concrete reference. It does not remove the need to inspect each artifact and license.

flowchart TD
    A["AI release"] --> B["Paper"]
    A --> C["Code"]
    A --> D["Weights"]
    A --> E["Data information and datasets"]
    A --> F["Demo, API, or product"]
    B --> G["Method visibility"]
    C --> H["Code license"]
    D --> I["Model terms"]
    E --> J["Provenance and data rights"]
    F --> K["Access and service terms"]
    H --> L["Accurate openness description"]
    I --> L
    J --> L
    K --> L

Why one label creates confusion

An AI system is assembled from pieces that can have different owners and terms. The inference code may use a permissive software license. The weights may use a custom model license. The training data may be unavailable. Evaluation data may prohibit commercial use. A hosted service may add another contract.

Calling that entire arrangement open because a GitHub repository exists hides the part that matters to the next user.

The reverse is also true. A system may not meet a strict open source definition and still provide valuable public research. Accurate language lets readers understand both the value and the limit.

Start with the freedoms

The OSI definition says an Open Source AI system must grant freedom to use, study, modify, and share it for any purpose. The phrase for any purpose is important. A license limited to academic or noncommercial research does not grant that same freedom.

The definition also addresses the preferred form for making modifications. It calls for detailed data information, the complete code used to train and run the system, and model parameters under suitable terms.

The definition does not demand that every original training datum be redistributed. It does require enough information about the data and process for a skilled person to understand and recreate a substantially equivalent system using the same or similar data.

This is one current standard, not a universal legal ruling. A specific use may still require legal review.

Read the paper as evidence, not permission

A public paper can describe architecture, training, benchmarks, ablations, and limitations. It helps a reader study the method.

The paper does not automatically grant permission to use the authors' code, weights, datasets, trademarks, or hosted implementation. Copyright access to the paper is also separate from rights in the underlying artifacts.

Movie Gen makes this distinction visible. Meta published a detailed Movie Gen research record and selected outputs. In the reviewed 2024 sources, the original models were not offered as an unrestricted public download. Public research did not equal an open model release.

Inspect code and weights separately

Code can reveal preprocessing, model architecture, training loops, inference, and evaluation. Its license controls how the code may be used and redistributed.

Weights capture learned parameters. Their terms may differ from the code. A release can make weights downloadable while limiting commercial use, certain domains, redistribution, or derivative models.

Spirit LM is a useful example. Its official repository exposes weights and inference code. The FAIR Noncommercial Research License limits covered materials and outputs to noncommercial research uses. “Public research model under a noncommercial license” is accurate. “Open source model” is too broad under the OSI use-any-purpose test.

Data visibility is not dataset reuse

A model card may describe data categories and processing without exposing the dataset. A paper may publish aggregate statistics without enough provenance to reproduce training. A dataset may be downloadable but carry its own limits.

Record data information and dataset access as two separate fields. Then check collection rights, consent, privacy, labeling, filtering, geographic scope, deletion, and downstream use.

This is especially important when a model processes voices, faces, personal images, medical records, private documents, or physical environments.

A hosted service is another layer

A demo shows that someone can interact with a system. An API provides a programmatic surface. A product adds terms, access controls, regions, pricing, retention, support, and change risk.

None of those layers proves access to the underlying code or weights. A hosted product can be useful while remaining closed. An open artifact can exist without any hosted service.

The correct description names both. “A hosted service built from a closed model” and “downloadable research weights without support” tell readers much more than open or closed alone.

Maintenance changes the practical answer

Licensing describes permission. Maintenance describes present operating risk.

The Sparsh repository exposes code and checkpoints under noncommercial terms, but it was archived on April 1, 2026. A team may still inspect the work. It should also expect to own dependency updates, security review, bug fixes, and compatibility.

That later archive state does not rewrite what was available in 2024. It changes the answer for a team evaluating the artifact now.

Use precise language

Write public paper when the paper is public. Write source-available when code can be inspected but the license does not meet the intended open source standard. Write open weights only when the parameters and their terms are clear. Write research license when use is limited to research. Write limited pilot when only selected participants had access. Write hosted product when users can access a current service under current terms.

[[Meta AI Release Map CoTracker3 Movie Gen Spirit LM and Sparsh]] applies that vocabulary to the four E044 projects. [[How to Evaluate an AI Research Release]] then tests whether the available artifact is useful for a real decision.

Editorial note

This explainer was developed with AI assistance from the E044 source record, OSI's Open Source AI Definition 1.0, and the linked project artifacts and licenses. Dalton Anderson remains the author. Standards, license, legal, source, current-state, and founder review are mandatory before publication. This page is not legal advice. Publication is not authorized.