I am often asked about the difference between open-source AI models and open-weight models. The terms are frequently used as if they meant the same thing. They do not.

The distinction is more than technical terminology. It matters because education needs to decide not only which AI systems it can use, but also which systems it can inspect, adapt, govern and trust. A university, college or school choosing an AI model is making a decision about privacy, cost, infrastructure, local knowledge, research freedom and dependence on commercial providers.

The short version is this: open weights give you the trained model; open-source AI aims to give you the means to understand, modify and share the system.

What are open weights?

The weights of an AI model are the very large set of learned numerical parameters produced during training. They determine, together with the model architecture and inference software, how the model responds to an input.

When a company releases an open-weight model, it makes these final parameters available for download or use. In practical terms, this may allow an institution to run the model on its own servers, adapt it to a particular task, or build an application without sending every prompt to a commercial cloud service.

However, open weights normally do not include the complete training data, the code used to curate and process that data, the full training code, intermediate checkpoints or all the information needed to reproduce the model. The weights reveal the final state of the system, but not the complete history of how that state was produced.

The Open Source Initiative therefore describes open weights as an important step towards greater openness, but not the same thing as open-source AI. The organisation’s explanation is direct: without the training data or training code, outsiders cannot fully scrutinise or recreate the training process [1] [2].

What is open-source AI?

The Open Source Initiative’s Open Source AI Definition applies the familiar four freedoms of open source: users should be able to use, study, modify and share the system [1]. For an AI model, those freedoms require more than access to the final weights.

The preferred form for modification should include detailed information about the training data, including its provenance, scope, characteristics, selection and filtering. It should include the complete code used to process the data, train and evaluate the model, run inference and define the architecture. It should also include the model parameters themselves.

This does not mean that every byte of every training dataset can always be published. Personal data, copyrighted material and confidential information may not legally be distributable. In such cases, a meaningful open-source approach should still provide detailed information about what was used, how it was obtained and how it was processed.

The distinction can be summarised as follows:

The categories are not always perfectly tidy. Some projects are more open than others, and licences may impose conditions that complicate reuse. “Open” should therefore be treated as something to verify, not as a marketing adjective to accept.

Question	Open-weight model	Open-source AI model
Are the final weights available?	Usually yes	Yes
Can the model be run locally?	Often, depending on licence and hardware	Yes, subject to licence and infrastructure
Is the training code available?	Usually no or only partly	Yes, in the preferred form for modification
Is the training data available?	Usually no; its composition may be unclear	Available where legally possible, with detailed data information where it is not
Can the training process be reproduced?	Usually not fully	The aim is to make meaningful reproduction possible
Can the community inspect and improve the system?	Mainly through fine-tuning and application work	Across data, code, training, evaluation and model components
Does the label guarantee safety or quality?	No	No

Education has particular reasons to care about openness. Schools, colleges and universities work with children, students, staff, research participants and sensitive institutional information. They also have public responsibilities that go beyond maximising efficiency or shareholder value.

Why does the distinction matter in education?

A completely closed AI service may be convenient, powerful and easy to access. But an institution may not know what data were used to train it, how its outputs are produced, what information is retained, how its behaviour changes after an update or whether its assumptions reflect the language and experience of the learners it serves.

Open weights can improve the situation. They can support local deployment, private experimentation and institutional control. Full open-source AI can go further by supporting research, audit, adaptation and public scrutiny. Neither approach removes the need for governance, but both can reduce dependence on a single remote provider.

The educational advantages of open-source AI

The strongest educational advantages of open-source AI are transparency and inspectability. Students and researchers can study how a system is assembled, what data and methods shaped it, how it is evaluated and where its limitations lie. This can turn AI from a mysterious consumer service into an object of study.

That matters for AI literacy. Students should not only learn how to write prompts. They should also learn how models are trained, how data selection influences outputs, how evaluation works, why models fail and how technical design is connected to social and political choices. Open systems make that kind of learning more feasible.

A second advantage is adaptation to local needs. A university or public education system may want a model that works well in a minority language, a regional curriculum, a particular professional field or a culturally specific context. An open development ecosystem makes it easier for local researchers and educators to contribute rather than waiting for a global provider to decide that their needs are commercially attractive.

A third advantage is collaborative research. Open code, data information, evaluation tools and model checkpoints can enable researchers to reproduce results, identify weaknesses and propose improvements. This is especially important in education, where the relevant question is rarely whether a model can produce fluent text. The question is whether its use improves understanding, agency, inclusion, judgement or access.

A fourth advantage is digital sovereignty. The United Nations University argues that open-source AI can help universities, public bodies and countries move from merely consuming AI to developing it, provided that openness is accompanied by investment in skills, infrastructure, research capacity and governance [5]. For education systems, this means maintaining the ability to make choices rather than becoming permanently dependent on a small number of foreign technology companies.

Finally, open-source AI can support the creation of public-interest infrastructure. UNESCO has linked open AI models with open data, open software and open educational resources as possible digital public goods when they comply with relevant law, avoid harm and contribute to sustainable development [4]. This suggests a different model of educational technology: not simply a product purchased by each institution, but shared infrastructure that can be improved and governed by communities.

The educational advantages of open-weight models

Open-weight models may be more immediately practical for many education institutions because they are easier to obtain and deploy than fully open systems. A college or university may not have the capacity to reproduce a large model’s training process, but it may have the technical ability to download a model, run it locally and connect it to selected institutional resources.

This can support privacy. Student records, draft assessments, research documents or internal materials need not automatically be sent to an external provider. Local deployment can be especially valuable in libraries, research services, counselling contexts and institutions working with sensitive data. The Ontario Council of University Libraries has highlighted local open-weight models as a way to explore AI while aligning more closely with professional commitments to privacy, intellectual freedom and equitable access [6].

Open weights can also reduce vendor lock-in. An institution can choose its own model engine, hosting arrangement and application layer. It may be able to move between models, fine-tune a system or maintain a service when an external provider changes its prices or terms.

There may be cost advantages, particularly where an institution has repeated, high-volume or predictable workloads. Instead of paying for every request through an external API, an institution can run a model on its own infrastructure. But “free to download” does not mean free to operate. Hardware, electricity, storage, technical expertise, cybersecurity, updates, monitoring and staff time all have costs. A local model may also be slower or less capable than a commercial frontier model.

Open weights can be useful for teaching and experimentation. Students can compare models, test how fine-tuning changes behaviour, examine errors, explore quantisation and learn about deployment. They can build applications that would be difficult to develop if every experiment required a paid cloud service and the approval of an external provider.

What neither approach solves automatically

It would be a mistake to treat open-source AI as inherently educational or open-weight AI as inherently trustworthy. Openness is valuable, but it is not a substitute for quality, safety or pedagogy.

An open model may reproduce bias, hallucinate, generate harmful content or perform poorly in a particular language. It may have been trained on data whose provenance is unclear. Its licence may limit certain uses. Its documentation may be incomplete. A model that can be inspected may still be too large or complex for an education institution to audit meaningfully.

There are also security and misuse risks. Open models can be adapted for harmful purposes, and publishing code and weights can make it easier for bad actors to remove safeguards. Open projects may lack dedicated technical support or guaranteed security updates. Responsibility does not disappear when a model becomes downloadable; it moves towards the institutions and communities that deploy it.

UNESCO’s discussion of openness makes a further point. Many systems described as open are open in only a limited sense. The term open-washing describes the presentation of a proprietary system as open while crucial aspects of its data, training, weights or governance remain controlled by the provider [3]. Education institutions should therefore ask precisely what is open rather than relying on the model’s brand or publicity.

A practical checklist for education institutions

Before adopting an open-source or open-weight model, an institution should ask:

1. What exactly is available? Are the weights, architecture, inference code, training code, data information, evaluation data and documentation provided?
2. What does the licence permit? Can the institution use, modify, share and fine-tune the model for its intended educational purpose?
3. Where will the system run? Can it operate locally or in a trusted institutional environment, and is the hardware adequate?
4. What happens to user data? Are prompts, uploaded documents, student work or interaction histories retained, reused or transmitted elsewhere?
5. How will the system be evaluated? Has it been tested for accuracy, bias, language performance, accessibility, safety and educational usefulness in the relevant context?
6. Who maintains it? Is there a clear process for updates, security patches, incident reporting and withdrawal if the model becomes unsafe?
7. How does it support learning? Does it promote explanation, questioning, feedback and learner agency, or does it simply produce answers?
8. Can the institution leave? Can data, configurations and applications be moved if a provider or model community disappears?
9. Who has a voice in governance? Are teachers, students, technical staff, researchers and affected communities involved in decisions?

These questions are more useful than asking whether a model is simply “open”.

Open for education, not just open for developers

The deeper issue is not whether education should choose open source or open weights as if one were always superior. The two approaches solve different problems.

Open weights offer a practical route to local deployment, privacy, experimentation and reduced dependence on external APIs. They may be the most realistic starting point for many institutions. Open-source AI offers a stronger basis for reproducibility, public scrutiny, local adaptation and collaborative improvement. It is especially important for research and for institutions that want to build long-term capability rather than merely consume a model.

The choice may therefore be layered. A university could use an open-weight model for a private pilot while supporting open-source research, contributing evaluation data and demanding greater transparency from providers. A school or college might use a locally hosted open-weight model for a limited, low-risk purpose while relying on shared open educational resources and open standards. A national education system might invest in public infrastructure that allows institutions to share models, tools, documentation and safeguards.

The goal should not be openness for its own sake. It should be educational agency: the ability of educators, learners and public institutions to understand the systems they use, shape them for their own purposes, question their outputs and change providers when necessary.

That is why the difference between open source and open weights matters. Open weights can give education access to a model. Open-source AI can help education develop the knowledge, control and collective capacity to decide what AI should do.

References

[1] Open Source Initiative. (n.d.). The Open Source AI Definition—1.0. Retrieved September 16, 2026, from https://opensource.org/ai/open-source-ai-definition

[2] Open Source Initiative. (n.d.). Open weights: Not quite what you’ve been told. Retrieved September 16, 2026, from https://opensource.org/ai/open-weights

[3] Amiel, T., Cox, G., & de la Higuera, C. (2025). Open education principles: Resisting the metrics of AI black boxes. UNESCO. https://www.unesco.org/en/articles/open-education-principles-resisting-metrics-ai-black-boxes

[4] UNESCO. (2026, April 7). UNESCO’s open solutions included in Digital Public Goods Alliance roadmap to advance inclusive knowledge ecosystems. https://www.unesco.org/en/articles/unescos-open-solutions-included-digital-public-goods-alliance-roadmap-advance-inclusive-knowledge

[5] Marwala, T. (2026, June 25). Open-source AI and the choice before us. United Nations University. https://unu.edu/article/open-source-ai-and-the-choice-before-us

[6] Weaver, K. D. (2026, August 24). AI for academic libraries: Open-weight AI models for local and private use. Choice. https://www.choice360.org/libtech-insight/ai-for-academic-libraries-open-weight-ai-models-for-local-and-private-use/