Open-Weight vs Open-Source AI Models Explained

Closed, open-weight and open-source describe how much of an AI model its maker shares. A closed model stays on the maker's servers. An open-weight model lets you download the trained parameters, called weights, and run them yourself. An open-source AI system, by the Open Source Initiative's definition, goes further and also shares information about the training data and the code used to build and run it. The labels are often mixed up, so the details are worth knowing.

What are weights?

When an AI model is trained, it adjusts millions or billions of numbers. Those numbers are the weights, and they hold what the model has learned. Releasing them is like releasing the finished product without the full recipe and ingredients. With the weights and suitable software you can run the model, and often tune it further, without sending your data to its maker.

Closed models

In a PBS NewsHour explainer, a researcher describes closed AI as the case where the developer keeps the model's internals secret and treats the model as private property. Users reach it through an app or a service, and cannot inspect or change how it works.

Open-weight models

The same PBS explainer says that open-weight developers release the trained weights, often with less restrictive reuse terms than a closed model, and names DeepSeek and Alibaba's Qwen as examples. It adds that the community has adopted such models quickly, though many developers do not consider them fully open source.

Open weight says nothing about the licence. Some open-weight releases allow almost any use, while others add conditions. The PBS article notes that Meta's LLaMA, released in February 2023 with its inference code and weights, is described by the Open Source Initiative as not truly open source, because its licence prohibits commercial reuse.

Open-source AI: a formal definition

The term open source comes from the free software movement of the 1980s and 1990s, which argued for four freedoms: to run, study, modify and share software. The Open Source Initiative (OSI) applies these to AI in its Open Source AI Definition, version 1.0. It says an open-source AI system must let people:

• Use the system for any purpose without asking permission.
• Study how it works and inspect its components.
• Modify it for any purpose, including changing its outputs.
• Share it with others, modified or not, for any use.

To make those freedoms real, the definition says people need the preferred form for making changes. For an AI system, the OSI lists three kinds of material:

• Data information: enough detail about the training data for a skilled person to recreate a comparable system, including where the data came from, how it was selected and processed, and where public or third-party data can be obtained.
• Code: the complete source used to train and run the system, including data processing and filtering, training settings, validation, inference and the model architecture.
• Parameters: the model weights or other settings, such as key intermediate checkpoints and the final optimizer state.

Notice what is and is not demanded. The OSI asks for information about the training data, not necessarily the data itself. The PBS explainer says some developers question whether datasets of that size can practically be distributed. The OSI page excerpted here does not give a release date, so check the OSI site for the current status of the definition.

Why the distinction matters

• Privacy and control: running a downloaded model on your own hardware means your prompts need not leave your machine.
• Cost and flexibility: with weights you can often adapt a model to your own task.
• Legal clarity: the licence decides whether you may use a model commercially, redistribute it or build on it.
• Scientific checking: sharing code and data information lets others test how a model was made.

The terms are contested

Not everyone agrees on where to draw the line. Companies sometimes call a model open source when it is open weight, and developers sometimes argue that a strict definition makes little sense when training datasets are huge or contain material that cannot be shared. Treat any label as a prompt to read the licence and the release notes. This article is general information and does not give legal advice.

Summary

• Closed: internals kept private, used through the maker's service.
• Open weight: trained weights are downloadable, licence terms vary.
• Open source (OSI definition): weights, code and data information are shared under terms that allow use, study, modification and sharing.

Sources

• The Open Source AI Definition 1.0, Open Source Initiative
• What's the difference between closed, open-source and open-weight AI, PBS News