You use proprietary models through the provider, while open models can be run yourself; what matters for your business is control over data, running costs, quality for your tasks and the effort of maintenance. Many businesses do best with a mix: a proprietary model for demanding tasks, an open one for sensitive or frequent routine work.
What is the difference between open and proprietary models?
With proprietary models you reach the provider's model through an application or interface. The parameters stay with the provider. With open-weight models you download the parameters and run the model on your own hardware or at a host of your choice.
The term "open source" is disputed for AI. The Open Source Initiative has published a definition of open source AI that requires freedoms to use, study, modify and share and includes information on training data and code. Many so-called open models publish only the weights and restrict use by licence. Read the licence before you deploy a model.
Which criteria decide?
The table compares the two routes.
| Criterion | Proprietary (through the provider) | Open weights, self-run |
|---|---|---|
| Control over data | Data goes to the provider, governed by contract | Data can stay in-house |
| Cost | Usage-based, low fixed costs | Fixed costs for hardware and operation, low cost per request |
| Performance | The most capable models are often proprietary | Often sufficient for clearly defined tasks |
| Effort | Low | Set-up, updates, monitoring, security |
| Support and availability | From the provider, sometimes with commitments | Falls to you or your contractor |
| Dependence | On provider, prices, terms | On your own skills and hardware |
| Flexibility | Limited adaptation | Adaptation and fine-tuning possible |
When does which route pay off?
Proprietary models pay off when you want to start quickly, need the best possible quality and the data situation allows use. You save operating effort and receive continuous improvements. Open models pay off when data must not leave the building, when very many similar requests arise or when you want to tune a model to your task.
A middle path is the hybrid: a proprietary model handles difficult, uncritical tasks, an open model on your own hardware handles the sensitive or frequent routine tasks. What self-hosting looks like is explained in Running AI models locally: when your own server beats the cloud.
How do you keep switching in your own hands?
Do not tie your processes firmly to one model. Wrap the call to the model behind an interface of your own, keep your instructions separate from the model and maintain a test set. Then any new model can be checked in a few hours, as described in How to evaluate a new AI model. Criteria for commercial providers are in Claude, ChatGPT or Gemini for business use.
Legally, the GDPR and the EU AI Act apply regardless of model type. An exemption for open models covers only parts of the obligations and does not release deployers from their own duties. This is not legal advice.
- Tasks sorted by sensitivity and frequency
- Licence terms of the candidates read
- Total costs of both routes calculated over a year, including operation
- Operating skills for self-hosting clarified
- Test set created and both routes compared
- Call to the model built to be swappable
Conclusion: not either-or
The choice between open and proprietary follows your data, your skills and your tasks. Check both routes with the same test set and build so that the model can be swapped. If you would like support, see our AI automation service or describe your requirements.
