Open-source AI

Also known as: open source AI

On this page 7

Open-source AI is an artificial intelligence system released under terms that grant anyone the freedom to use, study, modify and share it, together with the components needed to exercise those freedoms in practice: the model weights, the training and inference code, and sufficiently detailed information about the training data. Among the AI fundamentals, the concept matters because a trained model is not ordinary software, so the classic idea of publishing source code does not by itself make a model open.

The term is often used loosely. Many models described as open are in fact open-weight models, whose parameters can be downloaded but whose licenses restrict use or whose training data remain undisclosed. Telling these categories apart is essential for anyone who plans to build products, conduct research or meet regulatory duties with a published model.

What makes an AI system open source

An AI system is open source when its license and its published artifacts together allow the four freedoms known from free software: use for any purpose, study of how the system works, modification, and redistribution. The Open Source AI Definition, published by the Open Source Initiative in 2024, translates these freedoms to machine learning by naming the preferred form for making modifications to a model.

For a neural network, that preferred form has three parts. Parameters are the learned weights and configuration. Code covers the scripts for data processing, training, evaluation and inference. Data information describes the training data in enough detail that a skilled person could build a substantially equivalent system, even when the raw data itself cannot be redistributed for legal or privacy reasons.

The table below shows which components are required and why each one matters to a downstream user.

ComponentWhat is publishedWhy it matters
Model weightsTrained parameters and configuration filesEnable inference, fine-tuning and inspection
Training codeData pipelines, training scripts, hyperparametersAllows retraining and verification of claims
Inference codeModel architecture implementation and runtimeLets users run the model on their own hardware
Data informationSources, filtering, labeling and mixing methodsSupports audit of bias, provenance and copyright
LicenseTerms approved as open by a recognized bodyGuarantees use without field-of-use restrictions
Components of an open-source AI release

Degrees of openness in released models

Openness in AI is a spectrum rather than a binary property, and most published models sit somewhere between fully closed and fully open. The position depends on which artifacts are released and on the conditions attached to them.

Researchers often describe these levels with the following categories. The boundaries are not standardized, so a careful reader checks the actual license and release notes instead of relying on a label in a press announcement.

LevelTypical releaseMain restriction
Closed or API-onlyAccess through a hosted interfaceNo weights, no inspection of internals
Open-weight, restricted licenseDownloadable weights with usage policyUser caps, banned uses or noncommercial terms
Open-weight, permissive licenseWeights under Apache 2.0 or MITTraining data and code often withheld
Open-source AIWeights, code and data informationFew or none beyond attribution
Fully open scienceEverything above plus raw datasets and checkpointsLimited mainly by data rights
Common openness levels for AI models

A frequent source of confusion is the acceptable use policy attached to many popular open-weight models. Clauses that forbid certain applications or require a separate license above a user threshold are field-of-use restrictions, and under the Open Source AI Definition they disqualify a model from being called open source.

Open-source AI compared with proprietary models

The choice between open-source AI and a proprietary AI model is a trade-off between control and convenience. Open models can run on local infrastructure and be adapted freely, while proprietary models usually offer stronger out-of-the-box capability, managed hosting and vendor support.

CriterionOpen-source AIProprietary AI model
Data controlRuns on premises or in a private cloudPrompts and data pass through the vendor
CustomizationFull fine-tuning and architecture changesLimited to vendor tuning options
TransparencyWeights and methods can be auditedInternals and data largely undisclosed
Cost structureHardware and engineering effortPer-token or subscription fees
Capability ceilingOften behind the frontier by monthsTypically leads on frontier benchmarks
Vendor riskNo lock-in, version is permanentModel changes or retirement by vendor
Open-source and proprietary AI side by side

The capability gap has narrowed considerably since 2023. For many enterprise tasks such as classification, extraction and retrieval-augmented question answering, a well-tuned open model of moderate size performs comparably to a larger closed one, although the most demanding reasoning tasks still tend to favor frontier AI systems.

Why organizations choose open models

Organizations adopt open-source AI mainly for data sovereignty, reproducibility and long-term stability. A hospital, bank or public agency may be unable to send sensitive records to an external service, and a self-hosted model removes that obstacle.

  • Research reproducibility: published weights and code let independent teams verify results and study model behavior.
  • Domain adaptation: techniques such as LoRA fine-tuning adapt a foundation model to legal, medical or industrial vocabulary.
  • Edge deployment: quantized models run on laptops, phones and embedded devices without network access.
  • Security review: auditors can inspect a model for backdoors, memorized secrets or unsafe behavior.
  • Cost predictability: inference costs depend on owned hardware rather than on changing vendor prices.

The decision table below summarizes when an open model is usually the better choice.

SituationRecommended approachReason
Regulated personal dataSelf-hosted open modelData never leaves controlled systems
Rapid prototype, small teamHosted proprietary APILowest setup effort
High-volume narrow taskFine-tuned small open modelLower cost per request at scale
Academic study of model internalsFully open releaseRequires weights, code and data details
Hardest reasoning workloadsFrontier proprietary modelHighest current capability
Choosing between open and closed models

Risks, limitations and common misconceptions

Open release brings dual-use risk: once weights are public, safety fine-tuning can be removed and the model cannot be recalled. Policy debates about open-source AI therefore focus on whether the marginal risk of a release exceeds that of information already available elsewhere, a question on which experts disagree.

Operating an open model also shifts responsibility to the deployer. Security patches, evaluation, content filtering and compliance become in-house tasks, and the total cost of ownership can exceed API fees for low-volume use.

  • Misconception: open weights mean open source. Without an open license and data information, a model is open-weight only.
  • Misconception: open means free of obligations. Licenses still require attribution, and the EU AI Act keeps some duties for general-purpose models with systemic risk.
  • Misconception: open models are always less safe. Public scrutiny can expose flaws faster than closed review.
  • Misconception: data information equals the dataset. The definition accepts detailed descriptions when raw data cannot legally be shared.

The simple explanation

Think of an AI model as a finished cake. A closed model lets you buy a slice but never see the kitchen. An open-weight model gives you the whole cake to take home, yet keeps the recipe secret. Open-source AI gives you the cake, the recipe and a list of the ingredients, so you can bake it again, change it and share your version with others.

In everyday terms, open-source AI means nobody can take the tool away from you or forbid the way you use it, and anyone curious can check how it was made.

How it is used in practice

A regional hospital network wants to summarize discharge letters automatically, but its data protection rules forbid sending patient records to an outside service. The IT team downloads an open model with about 8 billion parameters released under the Apache 2.0 license and runs it on two on-premises GPU servers.

Clinicians annotate around 3,000 anonymized letters, and the team applies LoRA fine-tuning to teach the model local terminology and the required summary format. Because the training code and data documentation are public, the hospital’s auditors can review how the base model was built and record it in their risk register.

After a supervised pilot, physicians accept most summaries with minor edits, and no patient data ever leaves the hospital network. The hospital also keeps the exact model version indefinitely, which avoids unexpected behavior changes during clinical validation.