Does the EU AI Act apply to open-source AI? Exemptions and their limits
Open-source AI is not automatically outside the EU AI Act. This explainer walks through the Article 2 carve-out, the Article 53(2) exemption for free models, and the cases where either is lost.
Open source is not a pass. The EU AI Act open-source exemption is real, but conditional: it relieves specific documentation duties in specific situations, and it evaporates the moment a model is placed on the market as a high-risk system, classified as carrying systemic risk, or used for a prohibited practice. This explainer sets out what the free and open-source carve-outs actually cover, where each boundary sits, and what you still owe once they are stripped away.
Does the AI Act apply to open source models?
Usually yes, in part. Openness is a licensing choice, not a risk category. The Act never asks whether your repository is public; it asks what the system does, who placed it on the market, and in what role.
- What it does. Prohibited practices are unlawful regardless of licence, and Annex III high-risk uses bind whoever markets the system as high-risk.
- Your role. Provider, deployer, importer or distributor each carry different duties. Publishing weights may make you a provider; fine-tuning and selling the result under your own name certainly does.
- How it was released. Genuinely free and open-source releases earn conditional relief from certain documentation, copyright-policy and transparency paperwork.
The licence changes how much paperwork you owe. It does not change whether the Act reaches you.
The open source AI Act Article 2 carve out
The scope article contains a genuine free and open-source software AI act exception: AI components released as free and open source, with parameters, weights and usage information publicly available, are carved out of most high-risk requirements — unless they are placed on the market or put into service as high-risk systems, or as systems covered by the transparency article (chatbot disclosure, synthetic-content marking, deepfake disclosure).
Three conditions must hold:
- Genuinely open. The licence must permit access, usage, modification and distribution. Source-available or research-only terms will not do.
- Weights really public. Parameters, architecture and usage information all published; withheld weights break the condition instantly.
- Not marketed as high-risk. Once the component is commercialised as a high-risk or transparency-covered system, the carve-out disappears and the full regime attaches.
When it goes, it goes completely: technical documentation, a quality-management system and conformity assessment all follow the component down the value chain.
The AI Act free open source model exemption for general-purpose AI
General-purpose AI models get parallel relief. A GPAI model released under a free and open-source licence, with weights, architecture and usage information made public, is exempted from parts of the provider documentation package — most importantly the Union copyright-policy duty and the obligation to publish a detailed summary of training content, plus elements of the downstream information package.
What it never covers: systemic-risk obligations (models above the 10^25 FLOP threshold keep the full risk-mitigation, incident-reporting and cybersecurity duties), prohibited practices, and system-level transparency — users must still be told they are interacting with AI, and synthetic content still needs marking.
A note on citations: the exact enumeration of switched-off points sits in the GPAI obligations article and its exemption paragraph, and guidance is still bedding in, so confirm the current text with counsel before relying on it. The sound planning rule for open-source LLM compliance obligations is simple: the exemption trims paperwork, never accountability.
Four ways an open-source exemption is lost
- Placed on the market as high-risk. The big one. Wrap an open component into recruitment, credit, biometric or critical-infrastructure software — by you or by a downstream integrator — and the high-risk regime attaches: technical documentation, logging, human oversight, conformity assessment.
- Classified as systemic risk. Openness does nothing above the 10^25 FLOP threshold; the systemic-risk duties apply either way.
- Used for a prohibited purpose. Social scoring, untargeted facial scraping, prohibited workplace emotion inference — no licence excuses them. Fines reach €35 million or 7% of global annual turnover.
- Conditions quietly broken. Weights withdrawn behind an API, a licence swap, or a token “open” artefact can each destroy the status you relied on.
Record the licence, the release artefacts and the classification reasoning — a register row reading “open source, exempt” survives no inspection.
When fine-tuning or hosting makes you the provider
This is where the open source ai provider obligations eu catch mid-market teams off guard. Using an open model is not being its provider. Three activities make you one:
- Branding it. Ship a fine-tuned model or wrapper under your own trade mark and you present yourself to the market as the provider.
- Substantially modifying it. A fine-tune that changes capabilities, safety behaviour or intended purpose meaningfully makes the modifier the provider of the modified system.
- Retargeting it into a listed use. Tune a general assistant to rank job applications and you have moved it into Annex III territory.
Hosting sits on a finer line: renting compute differs from operating the model as a product with your own interface, terms and features. If your offering behaves like a product, expect it to be treated as one.
Nordwerk Analytics, a fictional 400-person German analytics firm, shows the pattern: it fine-tuned an open-weight model to score insurance-claims fraud, sold the output under its own name, and became the provider of a high-risk system — while the upstream lab stayed outside. Every fine-tune, fork and wrapper therefore needs an intake decision before production: who is the provider, which risk tier applies, which duties follow.
Deployers never inherit the exemption
The free and open-source relief runs to providers. A company using a system in its own operations gets nothing: deployer obligations apply in full whether the weights are open or closed.
- Human oversight. Competent named people with authority to intervene or stop the system. “We cannot inspect the weights” is not an oversight plan.
- Input-data relevance. Data fed to the system must be relevant and sufficiently representative for its purpose.
- Logging. Automatic recording of operations for at least the retention floor in the Act.
- Worker and user notification. People affected must be told they are subject to AI-assisted decision-making.
- AI literacy. In force since 2 February 2025, regardless of model licence.
Deployers in public services, credit and insurance also face the fundamental-rights impact assessment, and GDPR duties layer on unchanged. The division of labour in one line: the upstream lab may owe little because it released open weights; you, the EU deployer, owe oversight, logging, data quality and notification from day one.
Scenario table: role, exemptions and remaining duties
| Scenario | Role | Exemptions | Remaining duties |
|---|---|---|---|
| Lab releases open weights, not marketed as high-risk | Component provider | Scope carve-out; GPAI documentation relief | Prohibited and transparency-covered uses never exempt |
| Open model above the systemic-risk threshold | GPAI provider, systemic risk | None of substance | Risk mitigation, incident reporting, cybersecurity, downstream information |
| Fine-tuned model shipped under your own name as fraud scoring | High-risk provider | None — branding and modification defeat it | Technical documentation, quality management, logging, conformity assessment |
| Open-weight model triaging insurance claims in production | Deployer | None for deployers | Oversight, data relevance, logging, worker notification, FRIA, AI literacy |
| Self-hosted inference of an open model sold as a product | Infrastructure party or provider, on substance | Possible if purely infrastructure | Assess early — interface, terms and features decide it |
| Internal open chatbot for drafting, no listed high-risk use | Deployer, minimal or limited risk | No high-risk regime applies | AI literacy, interaction transparency, GDPR |
Commercial deployments crowd the bottom rows, where the exemption column reads “none”. At Valmara Retail Group, an open model used for staff scheduling carries the same deployer duties as any closed tool.
Releasing model weights under the AI Act: a pre-release checklist
- Choose a genuinely open licence granting access, usage, modification and distribution, and record the SPDX identifier.
- Publish weights, architecture and usage information together — a model card plus artefacts; an API-only “open” release does not qualify.
- Document the use-case envelope, stating that the release is not positioned for Annex III high-risk or transparency-covered uses.
- Check training compute against the systemic-risk threshold before release, not after.
- Run a short prohibited-use screen and file it with the release record.
- Brief downstream builders on disclosure and marking duties for chatbots and synthetic content.
- File the evidence: licence text, commit hash, artefact hashes and classification rationale in your AI register.
The Head of Legal owns the classification reasoning, the CISO owns release hygiene, the DPO confirms where personal data touches training or evaluation. Ninety minutes across three people now avoids a retro-fit under audit pressure later.
How to check your own position in three minutes
Renomira’s free exposure scan asks twelve questions about your AI usage and governance, then returns an exposure score, the ranked exposures with the exact article that fired on each, and the first three moves mapped to the dates that apply to you. A confirming email unlocks the full report, so the findings stay with your team rather than in a browser tab.
Frequently asked questions
Is releasing model weights enough to be exempt under the AI Act?
No. Weights must sit alongside a genuinely open licence and public architecture and usage information, and even then the relief evaporates if the model or component is placed on the market as high-risk, classified as carrying systemic risk, or used for a prohibited practice. Publication is necessary for the exemption; it is never sufficient on its own.
Does the free and open-source software AI Act exception cover fine-tunes?
No. That exception attaches to components and models released as open source, not to derived products. A fine-tune shipped under your own branding, or substantially modified, is assessed on its own merits — and the modifier becomes the provider of the result, inheriting whatever documentation, logging and conformity duties follow from its risk class.
Do deployers get any open-source exemption?
No. Deployer duties — human oversight, input-data relevance, logging, worker notification and AI literacy — apply to whoever puts a system to use, whatever licence the underlying model carries. Deployers are precisely the group the open-source exemptions never reach: the relief is written for providers of models and components, not for users.
What happens if a downstream customer turns our open model into a high-risk product?
The customer becomes the provider of that high-risk system and carries the heavy duties: technical documentation, quality-management system, logging and conformity assessment. Your own release-level position turns on how you published it — a genuinely open component not placed on the market as high-risk can still sit inside the scope carve-out. Keep the licence text, artefact hashes and stated intended uses; they are your evidence.
Which dates should an open-source team actually track?
Three. Since 2 February 2025, prohibited practices are unlawful and the AI-literacy duty applies to every team touching AI. From 2 August 2026, general applicability brings enforcement powers, transparency duties and the penalty regime into effect. High-risk obligations for Annex III systems follow on 2 December 2027. General-purpose model duties phase in around the general applicability date with a longer runway for models already on the market, so check current guidance before relying on any relief.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.
Keep reading
Does the EU AI Act Apply to US Companies? Territorial Scope Explained
10 min
ExplainersEU AI Act for SMEs and Start-Ups: What Applies, What's Exempt, and the Fine Rule That Helps You
9 min
ExplainersThe Digital Omnibus Delay: Your December 2027 High-Risk AI Deadline Is Already Ticking
12 min