Open-source vs proprietary AI models

Side by side

When each applies

Choose open weights when data must stay in-house, auditors need direct access to the model, or you want deep fine-tuning without vendor lock-in. You need a capable team to host, secure, monitor and update it.

Choose a proprietary model when you want the vendor to handle uptime, safety updates and patches. It also fits when you need a warranty, an IP indemnity and a service-level agreement.

Read the license either way. "Open" is used loosely. The Open Source Initiative's definition requires the code, the parameters and detailed training-data information, under terms that let anyone use, study, modify and share. Many models marketed as open publish only weights, under custom terms. Those terms can bar commercial use, ban certain applications or require you to pass restrictions to your own customers.

The EU AI Act treats openness narrowly. Under Art. 2(12), the Act does not apply to AI systems released under free and open-source licenses. The exemption disappears for high-risk systems, prohibited practices and Art. 50 transparency cases. For general-purpose AI models, Art. 53(2) waives the technical documentation and downstream information duties. The weights, architecture and usage information must be public. The copyright policy and training summary still apply, and models with systemic risk get no waiver. See general-purpose AI model.

Access type never removes deployer duties. A deployer of either kind still assesses, tests in its own context, oversees and monitors the system.

Worked example

Brightwater Clinics, a fictional hospital group in the Netherlands, wants a model to summarize clinical notes. It compares a proprietary model reached through a vendor's API with an open-weights model it could host itself.

Patient data must not leave Brightwater's servers, and its auditors want to test the model directly. Brightwater also has a strong machine learning operations team. Those facts point to open weights.

Before choosing, legal reads the license. It allows commercial use and fine-tuning but bans some medical diagnosis uses. Summarizing notes for clinicians to review is outside that ban, so the model is approved for that purpose only. The approved use goes into the AI inventory and the restriction into the acceptable use policy.

Brightwater fine-tunes the model on de-identified notes and red-teams it for leaks of patient data. It owns patching and monitoring. Had it chosen the API instead, it would have needed a data processing agreement, transfer checks and the vendor's evaluation evidence.

If Brightwater later sold the fine-tuned model under its own name, it would become a provider. See provider vs deployer and RAG vs fine-tuning.

The exam trap

The classic trap is "open means free of strings." Downloadable weights do not grant the right to use a model commercially or for any purpose. The license decides.

The tells:

• Data residency, direct inspection and deep customization point to open weights.
• Vendor-managed safety, support and indemnity point to a proprietary model.
• "Closed models are inherently more secure" is an absolute, and usually wrong.
• "The vendor controls it, so we have no duties" is wrong. Deployers keep their duties.
• The AI Act's open-source exemption never covers high-risk or prohibited uses.

Under the Commission's GPAI guidelines, a downstream modifier can become the provider of the modified model. The indicative line is using more than a third of the original training compute. Drill the four model axes in lesson IV.A.2 and license families in lesson IV.B.4.