Learn AI /AI Foundations /Open-weight model
Open-weight model
- Intermediate
- AI Foundations
- Policy officers
- Technical readers
Also called: open weights.
Back to the milestone you came from
In five seconds
What it is
A model whose trained numerical parameters are publicly released so others can run, inspect or adapt it, even if the training code, data or licence are not fully open.
In plain words
An open-weight model is a model whose trained settings have been shared so other people can run it themselves.
An everyday example
A photographer publishes a preset file containing hundreds of tuned settings. Anyone with compatible editing software can load it and, if the terms allow, alter it. The file does not include the source photographs, the experiments used to tune the settings or necessarily the editing software itself. An open-weight model is similar: the tuned numerical values are available, while much of the process and material that produced them may remain unavailable.
A scoring sheet for duplicate records is small and easy to read, while real model weights can contain millions or billions of numbers and are not directly meaningful to most people. Also, a human clerk can explain a judgment in ordinary language more easily than a model can. The analogy helps with the idea of sharing trained parameters, but it does not show the scale, complexity or unpredictability of model behaviour.
Remember thisOpen weights provide the trained parameter values, not necessarily the training data, code, full development record or unrestricted permission to use and adapt the model.
This term is contested
The boundary of this term is genuinely unsettled. Many people use 'open-weight' narrowly to mean that trained weights are downloadable, while others expect some combination of code, documentation or permissive rights as well. The term is useful, but it does not by itself settle how open the overall system really is.
Understand it
Understand it
Why it matters
The phrase affects procurement, risk assessment and governance. If weights are released, an organisation may be able to run the model itself, test it more deeply and adapt it to local needs. But that does not tell you whether the training data are known, whether the licence permits your use, or whether the system is safe, fair or suitable for sensitive work. Knowing the distinction helps avoid overclaiming transparency.
Open weights compared with fully open and fully closed release
After viewing this, a reader should be able to say that open weights means the trained parameters are released, but other parts of the system may still be closed or restricted.
01
Same system, different release choices
Show that an AI system has several parts, and organisations can share some parts without sharing all of them.
Render three vertical panels with identical rows: Weights, Code, Data, Training method, Licence. Use consistent icons and colour states: released, partial, not released.
02
Closed model
In a closed release, the public may get an interface or service, but not the internal weights or other building parts.
In the left panel, mark all five rows as not released except optionally a small footer tag saying 'public access via service only'.
03
Open-weight model
Here the weights are released, so others can run the trained model. Other rows remain mixed: some may be closed, partly described or restricted.
In the middle panel, mark Weights as released. Mark Code, Data and Training method as partial or not released. Mark Licence as either partial or restricted with a note icon.
04
More fully open release
A more open release may share code and documentation too, and sometimes data, but this is a broader condition than open weights alone.
In the right panel, mark Weights and Code as released, Training method as released, Data as partial or released, and Licence as permissive. Add a small note: 'varies by project'.
A three-column comparison. Each column shows the same five components: weights, source code, training data, training method and licence terms. Read left to right: fully closed release, open-weight release and more fully open release. Each component is marked as released, partly released or not released, making clear that open weights describes one component, not the whole AI system.
Hover or tap each row to reveal a short tooltip explaining that component and why its release matters for reuse, auditing or legal permissions.
Describe this without the diagram
This diagram compares three ways an AI system can be released. Each column contains the same five parts: weights, source code, training data, training method and licence terms. In the closed column, none of these parts are released to the public, though people may still use the system through an online service. In the open-weight column, the weights are released, which means others can run the trained model, but the code, data, method and licence may still be partly closed or restricted. In the more fully open column, the weights, code and method are released, and the data and licence are more open, though this varies by project. The main lesson is that 'open weights' refers only to release of trained parameters, not full openness of the whole system.
With reduced motion: Show the same three-column comparison as a static table with colour labels and a visible legend for released, partial and not released. Tooltips are replaced by short text under each row title.
An open-weight model is a machine-learning model whose weights have been released. Weights are the many numerical values that shape how the model turns an input into an output. In simple terms, they are the trained settings produced after learning from data.
What goes in depends on the model: text, images, audio or other data. What happens is that software loads the released weights into a model architecture, which is the design of the model, and then computes an output such as a summary, translation or prediction. What comes out is the model’s result, not a guarantee of truth.
The important point is what has been shared. If the weights are public, others can usually run the model on their own systems and often fine-tune it for new tasks. But open weights do not automatically mean the training data, source code, evaluation methods or legal rights are also open.
This is why the term is often confused with open-source AI, but they are not the same. Open source usually refers to software whose code is available under licences that allow use, study, modification and sharing. An open-weight model may release the weights while keeping other parts closed or restricting some uses.
How “A shared camera preset” maps onto the real thing
| In the story | In the AI system |
|---|---|
| the clerk’s scoring sheet | the released model weights |
| the duplicate-checking software in another office | software that loads and runs the model |
| past registry records used to train the clerk | training data that may remain unreleased |
| the full handbook for building the method | source code, training procedure and other artefacts that may stay closed |
A worked example
An information management officer at a regional hub tests a publicly released summarisation model in an isolated environment before an information and communications technology governance review.
Starting from: Prompt: "Summarise this public 220-word meeting note in three bullet points and keep all dates unchanged."
- The officer checks what was actually released: checkpoint files containing the trained weights, licence terms and enough format information to use a compatible runtime.
- The checkpoint is downloaded from the stated source and loaded into approved software on local infrastructure. The original training data and training pipeline are not part of the release.
- The software processes the meeting note using the learned parameter values in the checkpoint and generates a summary without sending the text to a hosted model service.
- The officer compares the output with the source, records an altered date and notes that access to the weights does not explain which training examples produced the error.
Result: The model can be executed locally because its weights are available in a usable format. Whether it can be fine-tuned, redistributed or used operationally still depends on the licence, supporting software, technical assessment and governance clearance.
In a United Nations workplace
A chief of section in a procurement unit is considering whether an open-weight model could help screen a long list of vendor applications against an approved roster profile before the file moves to substantive review.
Publicly released weights could be run and tested in an approved internal environment rather than through an external service. The unit could also adapt the model if the licence permits. It would still need compatible software, adequate infrastructure and evidence that the proposed use is reliable and lawful.
Before you try thisReal applications, personal data and procurement-sensitive submissions must not be loaded into a model or test environment that has not been formally approved for that data and workflow. Open weights do not establish security, fairness or fitness for purpose. They also do not justify automated eligibility or ranking decisions without authorised governance and meaningful human review.
Where you may hear it
- A vendor demo in a procurement meeting on local hosting
- An information and communications technology governance board reviewing architecture options
- A technical review before clearance of a model pilot
- An inter-agency working group comparing model adaptation options
Avoid confusion
Avoid confusion
Often confused with
Open-source AI
AI systems released under licence terms that let others inspect, use, modify and redistribute source code and other materials needed to run them.
Proprietary model
An AI model whose owner uses licence terms or technical controls to restrict access to or reuse of key components, such as its weights or code.
Foundation model
A machine-learning model pretrained at large scale on broad data can be adapted for many tasks without training a new model from scratch.
Side-by-side
Common misconceptions
Common misconception: If a model is open-weight, it is open source.
Correction: Not necessarily. Open weights means the trained parameters are available. Open source usually concerns software code and the legal rights to use, study, modify and share it. A project can release weights while keeping code, data or some usage rights restricted.
Common misconception: If weights are public, the model is transparent enough to audit easily.
Correction: Released weights can improve access for testing, but weights are huge numerical files, not a plain-language explanation. Meaningful auditing may still require documentation, evaluation results, system prompts, safety measures, training details and information about intended and prohibited uses.
Common misconception: Open-weight models are automatically safer because more people can inspect them.
Correction: Broader access can support testing and independent scrutiny, but it can also make powerful capabilities easier to redistribute or modify. Safety depends on the model, the safeguards, the deployment setting, the licence and the oversight process, not on the label alone.
Quick check
Which statement best describes an open-weight model?
The defining feature is release of the trained parameters. Other parts of the system may remain closed, partly described or legally restricted. Public availability also does not prove accuracy or safety.
Go deeper
Go deeper
The simple version
An open-weight model is a model whose trained settings have been shared so other people can run it themselves.
The technical version
An open-weight model is a trained machine-learning model for which the learned parameter values, usually checkpoint files containing weights, are publicly released for download and execution on compatible infrastructure. The release may permit inference, fine-tuning or further adaptation, depending on the licence and format. However, this does not imply release of the training dataset, data-cleaning pipeline, source code, architecture details beyond what can be inferred, optimiser state, evaluation suite, safety methods or unrestricted legal rights. In practice, 'open weight' describes the availability of a specific artefact in the model lifecycle, not complete openness of the whole system.
Reviewed external resources
Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
This UNESCO recommendation gives a governance and ethics frame for evaluating AI use and release choices, which suits policy and oversight readers.
Intermediatelink checked 2026-07-31
EncyclopaediaOpen weights (opens in a new tab)Wikipedia
This Wikipedia entry gives a general overview of the term and related ideas, which suits readers who want a quick orientation before going deeper.
Intermediatelink checked 2026-07-31
ArticleMachine Learning Glossary (opens in a new tab)Google
This Google glossary offers short plain-language definitions for machine-learning terms around models and training, which suits readers building vocabulary.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Open-weight model (opens in a new tab)YouTube search
This YouTube search results page helps readers find multiple video explanations and comparisons of open-weight models, which suits people who learn well by listening and watching.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Open-weight model (opens in a new tab)Reddit search
This Reddit search results page surfaces community discussions and disagreements about the term, which suits readers who want to see how the phrase is used informally in practice.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Open weights provide the trained parameter values, not necessarily the training data, code, full development record or unrestricted permission to use and adapt the model.