Open-source models: definition and framing
Definition
Open-source models are AI language models whose weights are openly available and which can be run in your own data centre or with a freely chosen provider. Unlike closed models, usable only through a vendor's interface, they can be run locally, adapted and inspected. Licence terms vary and determine what is allowed.
Why the Term Matters
For mid-market firms, open-source models are rarely about ideology and mostly about control over data, cost and dependency. Anyone who may not send sensitive data to an external service finds an option here. For neutral consulting it matters to name the effort of self-operation honestly. Open models are no free ride; they shift responsibility for operation, security and currency into your own house.
When It Is Relevant
- Sensitive data must not leave your own sphere of control.
- Dependency on a single vendor should be limited.
- Costs at high, steady usage volume should be predictable.
- The model should be allowed to specialise on your own data.
When Not
When usage is rare or fluctuating, running your own infrastructure costs more than a hosted service. And when no in-house competence for operation and security exists, an open model quickly becomes a risk. Then a hosted vendor with clear data-protection commitments is often the more pragmatic path.
Industries That Use It
- Banks and financial services: process sensitive data in your own audited environment rather than an external service.
- Research and development: adapt models to your own domain data without releasing content externally.
- Government and public bodies: secure data sovereignty and traceability for regulated procedures.
- Medical technology: keep analysis in a controlled environment where access and origin are documented.
Typical Use Cases
- A model runs in your own data centre so no data leaves the house.
- An open model is specialised on your own terminology and documents.
- At high sustained volume, costs are lowered versus a hosted service.
- A vendor comparison assesses open and closed models for the same case.
- A model runs against data-protection requirements in a sealed environment.
Decision Guide
| Signal | Open-source model fits | Alternative |
|---|---|---|
| Data must not leave the house | assess self-hosting | hosted service is ruled out |
| Usage rare or fluctuating | rather not | hosted vendor is cheaper |
| No operations competence in-house | clarify competence first | hosted service with commitments |
| High, steady volume | cost advantage possible | a pure interface gets expensive |
Example
A financial services firm may not send customer data to an external service. We assess an open model in its own data centre against a hosted vendor with a data-protection commitment and place effort, cost and risk side by side. Only then is the decision made. This describes the method, not an asserted client metric.
Service Connection
The open-versus-hosted trade-off belongs in the AI potential analysis and the AI readiness check; operation is guided by AI implementation.
Sources
This framing describes common selection practice. Licence terms of open models differ and must be checked case by case against the planned use.