Open vs. closed AI models: what the difference means

When people call an AI model “open,” they can mean very different things: downloadable weights, a licence anyone can use for any purpose, or a fully public recipe of weights, training code and a detailed account of the training data.

A “closed” model is usually simpler to describe. You reach it through the developer's own application or API, and the weights stay with the developer. But closed does not automatically mean less private, less customisable or more expensive. Those outcomes depend on the terms, the deployment and your workload.

This explainer separates the questions that the open-versus-closed label tends to blur. If you need a refresher on what weights are and how a model uses them, start with our guide to what a large language model is.

Three separate questions hiding in one label

It helps to split “open or closed?” into three questions you can answer independently:

1. Can you get the weights? If you can download them, you can run the model on infrastructure you choose. If you cannot, you use it through someone else's service.
2. What does the licence let you do? Downloadable weights can come with broad permissions or with conditions on use, naming, redistribution and scale.
3. Can you study how the model was built? Weights alone show the end state of training. Training code and data information show how that state was reached.

Many open-weight releases answer “yes” to the first question and “partly” to the others, which is why “open” on its own cannot settle a decision.

What the Open Source AI Definition requires

The Open Source Initiative (OSI), the body that maintains the Open Source Definition for software, published version 1.0 of the Open Source AI Definition (OSAID) to answer a specific question: what is the “preferred form” for modifying a machine-learning system?[1]

Its answer has two parts. First, an Open Source AI must be available under terms that grant the freedoms to use it for any purpose, study how it works, modify it and share it. Second, exercising those freedoms requires three elements:[1]

• Data information: enough detail about the training data that “a skilled person can build a substantially equivalent system,” including its provenance, selection, labelling and filtering, and where public or third-party data can be obtained.
• Code: the complete source code used to train and run the system, including data processing, training settings and inference code.
• Parameters: the weights and other configuration settings, made available under OSI-approved terms.

The definition does not require every training document to be published. Where data cannot legally be shared, such as personal medical records, the OSI's FAQ requires a detailed description instead, a compromise the FAQ acknowledges is debated.[2]

Open weights are a different thing

The OSI draws a clear line between Open Source AI and open weights. Its explainer describes open weights as the final trained parameters, shared without the training code or full data transparency.[3]

“…these weights reveal only a fraction of the information required for full accountability.”

— Open Source Initiative, Open Weights: not quite what you've been told.[3]

With weights alone, you can run and fine-tune a model, but you cannot reproduce its training or check how its data was assembled. In its validation exercise for the definition, volunteers found that Pythia, OLMo, Amber, CrystalCoder and T5 passed, while Llama 2, Grok, Phi-2 and Mixtral did not, because components were missing or their legal terms were incompatible.[2] The OSI describes these as results of the definitional process, not certifications.

Ai2's current Olmo pages illustrate what the fuller approach looks like: the developer links model weights alongside pretraining, mid-training and post-training datasets, and its training and evaluation code.[4]

Licences decide what “open” lets you do

Downloadable weights tell you where a model can run. The licence tells you what you are allowed to do with it. Two examples show how wide the range is.

A custom licence: Llama 3.1

Meta's Llama 3.1 Community License, dated 23 July 2024, grants a royalty-free licence to use, copy, distribute and modify the Llama materials.[5] It also attaches conditions. Among them:

• Anyone distributing the materials, or a product containing them, must include the licence and display “Built with Llama.” A distributed model built or improved with Llama materials or outputs must include “Llama” at the beginning of its name.
• Use must comply with Meta's Acceptable Use Policy, which the licence incorporates by reference.
• If a licensee's products had more than 700 million monthly active users in the month before the release date, it must request a separate licence, which Meta may grant at its discretion.

Few organisations will meet the 700-million threshold. The broader point is that the terms are Meta's own, can change between versions and include use restrictions. Use restrictions of this kind sit uneasily with the OSAID's requirement that a system can be used “for any purpose and without having to ask for permission.”[1] Read the licence for the exact version you deploy; do not assume one Llama licence covers all of them.

A standard open-source licence: Apache 2.0

The Apache License 2.0 is an established, OSI-approved software licence.[6] It grants a perpetual, worldwide, royalty-free copyright licence to reproduce, modify and distribute the work, and a patent licence from contributors. Its main conditions concern notices: include the licence, mark changed files and keep the attribution notices.

OpenAI released its gpt-oss-120b and gpt-oss-20b open-weight models under Apache 2.0 in August 2025.[7] A familiar licence like this gives legal teams a known document to review. An Apache 2.0 licence on the weights does not, by itself, make a release Open Source AI under the OSAID: the definition also asks for complete training code and data information.[1]

What “closed” means in practice

A closed model is usually offered as a service. You send inputs to the provider's application or API and receive outputs. The provider keeps the weights, decides when versions change or are retired, and sets the terms for data handling.

Those terms can be detailed, and they differ by product. OpenAI's API documentation, for example, says that since 1 March 2023 data sent to its API is not used to train its models unless the customer opts in.[8] The same page explains that abuse-monitoring logs, which can contain prompts and responses, are retained for up to 30 days by default, and that stricter retention controls require OpenAI's approval. It also lists optional data-residency regions with their own conditions.

With a closed model, then, privacy is a contractual and configuration question: you assess the provider's documentation and your agreement rather than controlling the machines.

Closed does not always mean fixed, either. Some providers offer customisation through their platform; OpenAI's documentation lists a fine-tuning endpoint, for instance.[8] What you cannot usually do is take the resulting model and run it elsewhere.

Five practical differences

Cost

The two approaches charge in different shapes. A hosted API typically bills by usage, so the provider carries the cost of hardware and operations. Self-hosting an open-weight model shifts that to you: compute, storage, engineering time, monitoring and updates.

Hardware requirements depend on the model. OpenAI said in August 2025 that gpt-oss-120b can run within 80 GB of memory and gpt-oss-20b within 16 GB, because the weights ship in a compressed format.[7] Those figures are the developer's own and apply to those models; they say nothing about throughput under your traffic. Open weights also do not mean you must self-host: OpenAI listed several third-party platforms offering gpt-oss at launch.[7]

Which route costs less depends on your volume, latency needs and staff. Model both before deciding.

Privacy and data location

Self-hosting gives you physical control: inputs need not leave infrastructure you manage. The US National Telecommunications and Information Administration (NTIA) named this among the benefits of open foundation models in its 2024 report.[9]

“…they enable users to leverage models without sharing data with third parties, increasing confidentiality and data protection.”

— NTIA, Dual-Use Foundation Models with Widely Available Model Weights Report (2024).[9]

Control is not the same as security, though. A self-hosted model inherits your access controls, logging and patching habits; a provider with suitable terms may suit some organisations better than a poorly run internal server.

Customisation

With weights, you can fine-tune within the licence and keep the adapted model; with a closed model, customisation is whatever the provider offers. First ask whether you need to change the model at all. Many tasks are better served by supplying the right documents at run time, the approach covered in our retrieval-augmented generation explainer, or by fitting key material into the context window.

Safety controls

A hosted provider can update safeguards, monitor misuse and withdraw a model. Once weights are published, those controls no longer sit with the original developer. OpenAI's gpt-oss announcement states the risk plainly.

“Once an open-weight model is released, adversaries may be able to fine-tune the model for malicious purposes.”

— OpenAI, Introducing gpt-oss (2025).[7]

OpenAI reported testing adversarially fine-tuned versions of gpt-oss-120b before release; that is the developer's own evaluation. NTIA framed the policy question in terms of marginal risk: what open weights add compared with closed models and other existing technologies. In 2024 it concluded that current evidence was not sufficient to determine either that restrictions on open-weight models were warranted or that they would never be appropriate, and it recommended that the US government actively monitor risks.[9]

For a deploying organisation, self-hosting moves safety work to you: filtering, abuse monitoring and incident response.

Governance and accountability

Openness changes what you can inspect. With open weights, auditors can test the model directly but cannot see how it was trained. With full Open Source AI, they can examine the training process too.[3] With a closed model, you rely on the provider's documentation and contractual commitments.

Regulation also treats openness differently. Under Article 53 of the EU AI Act, providers of general-purpose AI models released under a free and open-source licence, with their weights and architecture and usage information made public, are exempt from two technical-documentation duties. The exemption does not apply to models with systemic risk, and the copyright-policy and training-content summary duties still apply.[10] Our EU AI Act explainer covers the wider framework, and the training-data and copyright guide covers the separate question of what models may be trained on.

AI Chronicle

A worked example: choosing for client-intake summaries

This is a hypothetical scenario created for illustration. No product was tested, and no costs or results are claimed.

Suppose a ten-person immigration law practice wants a model to draft summaries of sensitive client intake notes. Two staff members are comfortable with servers; nobody has trained a model.

The practice writes down its constraints before looking at any model:

Data: client notes are confidential; check professional and data-protection rules.
Volume: modest and predictable.
Customisation: summaries must follow the practice's own template.
Operations: no dedicated ML staff.
Review: a lawyer checks every summary before use.

Then it walks the three routes:

• A closed API. Least operating work. The practice would need to confirm the provider's training use, retention and region terms meet its obligations, and whether stricter retention controls are available to it. The template can likely be handled through instructions and examples rather than fine-tuning.
• An open-weight model under a custom licence, self-hosted. Notes stay on the practice's infrastructure. The practice would need to confirm the licence and acceptable-use terms fit its use, and accept the work of securing, patching and monitoring the server.
• An Apache 2.0 open-weight model through a third-party host. Fewer licence questions, but the data again leaves the building, so the host's terms need the same review as a closed API.

In this hypothetical, the deciding constraint is the professional rule on client data, not model quality. If the practice's obligations permit a provider with suitable terms, the closed API is lighter. If not, self-hosting is the route, with its operating budget. Either way, the practice would then compare a shortlist of models on anonymised sample notes, with a lawyer judging each output.

A decision guide

Use this as a starting point, not a verdict. Most organisations will mix approaches.

Whatever you choose, test candidate models on your own examples. Published scores are a starting point at best; our guide to what AI benchmarks measure explains why.

Questions readers commonly ask

Is an open-weight model the same as open source?

Not under the OSI's definition. Open weights give you the trained parameters. Open Source AI also requires the complete training and inference code and detailed data information, under terms granting the freedoms to use, study, modify and share.[1][3]

Can I use any open-weight model commercially?

Only if its licence allows your use. Apache 2.0 permits commercial use with notice conditions.[6] Custom licences can add conditions, such as Llama 3.1's naming rules, acceptable-use policy and large-platform threshold.[5] Read the licence for the exact version.

Are open models less safe than closed ones?

They carry a different risk: once weights are public, safeguards can be removed through fine-tuning.[7] Whether that adds meaningful risk compared with closed models is a question NTIA said requires ongoing monitoring.[9] For your deployment, the safety work you take on matters more than the label.

Can a closed model be customised?

Often, within limits the provider sets, such as a fine-tuning service.[8] You generally cannot export the customised model.

The next step

You now have four questions for any model announcement: Can I get the weights? What does the licence allow? Can I see how it was trained? Who handles data, updates and misuse?

Take them into our guide to judging a new AI model release, which covers how to read launch claims before you commit time to testing.

Sources and evidence note

This explainer relies on the OSI's definition and supporting documents, licence texts, developer documentation, a US government report and the EU AI Act text published by the European Commission's AI Act Service Desk. Developer statements about their own models are attributed as such. The law-practice scenario and decision guide are original illustrative material. No model was tested, and no cost, performance or safety result is claimed. All sources were accessed 15 October 2026.

Related reading

• How reasoning models work, and what changes
• Claude vs. ChatGPT vs. Gemini: choosing an assistant