Summary Under the proposed Cloud and AI Development Act (CADA), publishing a model's weights does not by itself make the licence "open source." Article 2(25) of the proposal would define "open source licence" by reference to the Interoperable Europe Act (Regulation (EU) 2024/903), Article 2, point (12). A licence that restricts how the model may be used β€” for example, non-commercial-only terms or scale caps β€” would likely fail that legal test even when the weights are freely downloadable. As proposed, "open-weight" and "open source" are not the same thing.

Detail

CADA introduces a specific legal hook for the term "open source licence" that diverges from the colloquial industry use of the word. To work out whether an open-weight model counts as open source under the proposal, you start with Article 2(25) of the CADA proposal, which reads:

"β€˜open source licence’ means open source licence as defined in Article 2, point (12), of Regulation (EU) 2024/903."

Regulation (EU) 2024/903 is the Interoperable Europe Act. CADA does not write its own substantive definition; it borrows the Interoperable Europe Act's, which centres on a licence that allows software to be freely used, studied, modified and shared. That is the test that matters under CADA, as proposed.

The open-weight vs open-source distinction

"Open-weight" models are those whose parameters (weights) are publicly available. That is a statement about availability of the artefact, not about the legal freedoms attached to it. The licence may still impose restrictions such as:

  • Commercial-use prohibitions β€” the licence forbids using the model commercially;
  • Scale limits β€” use is capped at a number of users or a level of compute;
  • Acceptable-use policies (AUPs) β€” the licence bans specific tasks (for example certain generated content).

These terms can make the weights available while keeping the licence short of the freedoms required by the open source definition CADA points to. If a licence restricts the freedom to use the software (for example "non-commercial only"), it would likely not satisfy the open source test imported by Article 2(25), and the model would not be "open source" under CADA as proposed.

Why the classification matters

The distinction has knock-on effects. The proposal repeatedly treats open source as a lever for the European cloud and AI ecosystem: it would encourage public sector bodies and Union entities to reuse open source components (Article 41) and to publish reusable software under an open source licence via a catalogue connected to the EU Open Source Solutions Catalogue (Article 42). A model that does not meet CADA's open source definition would not benefit from provisions framed around open source.

Separately β€” and this is an AI Act point, not a CADA one β€” the AI Act grants certain relief to providers of general-purpose AI models released under free and open-source licences. CADA does not change those AI Act provisions; it simply uses a definition of "open source licence" anchored in the Interoperable Europe Act. Whether a given model qualifies for AI Act open-source relief is governed by the AI Act's own text, which sits in its general-purpose AI provisions (Articles 51–56).

Edge case: community licences with acceptable-use limits

Many models ship under "community licences" that are open-weight but carry acceptable-use policies β€” often prohibiting uses that violate rights or applicable law. However well-motivated, such restrictions are still limits on how the software may be used. Under a strict reading of the open source definition CADA imports, usage-based restrictions beyond ordinary open source conditions (such as attribution or copyleft) could cause the licence to fall outside the "open source licence" definition in Article 2(25). The safe assumption, as proposed, is that any field-of-use restriction puts the licence at risk of failing the test.

What this means for you

For CTOs, architects and SMEs evaluating models:

  1. Audit the licence, not the download. Check whether the licence restricts commercial use, scale or specific applications. If it does, treat the model as not open source under CADA, as proposed.
  2. Mind the AI Act relief separately. Open-source relief under the AI Act turns on the AI Act's own definition and conditions. Do not assume a "community" or "open-weight" tag secures it.
  3. Plan procurement around the definition. Provisions encouraging open source reuse (Articles 41 and 42) are framed around the open source definition in Article 2(25). Models with restrictive licences may not slot into those reuse pathways.
  4. Weigh long-term dependency. A provider who can unilaterally change or revoke a restrictive licence is a continuity risk; genuinely open source alternatives reduce that exposure.

Common misconceptions

  • "Open-weight means open source." No. Open-weight is about availability of parameters; open source is about the freedoms the licence grants. A model can be open-weight but not open source.
  • "Acceptable-use policies are compatible with open source." Field-of-use restrictions are limits on the freedom to use the software and can disqualify a licence under the definition CADA points to in Article 2(25), as proposed.
  • "Publicly available code is always open source." The licence terms decide. Without the freedoms to use, study, modify and share, it is not "open source" under CADA's referenced definition.

Official sources

Related

This is general information about a draft EU regulation, not legal advice.