EXAM / AI research & technology

Open models and proprietary AI: compare rights and operating costs

Available model weights, open-source status and a hosted service are different choices with different responsibilities.

Open lattice and closed computing enclosures

Downloading model weights does not automatically give unrestricted commercial rights or make the whole system open source. Read the actual licence and the material released with it. A proprietary hosted API offers a different arrangement: you buy access to a service and rely on its contractual terms, availability and change policy.

Compare a specific deployment

Table scrolls horizontally

QuestionSelf-hosted model with available weightsHosted proprietary API
Who operates it?Your team or chosen operatorProvider runs the model service
Which costs remain?Hardware/cloud, engineering, security and maintenanceUsage fees plus integration, evaluation and oversight
Which rights matter?Commercial use, modifications and redistribution in the licenceData terms, allowed uses, service changes and exit

Do not use “open” as a shortcut

The Open Source Initiative’s AI definition addresses freedoms to use, study, modify and share, along with required information and components. Available weights alone are not sufficient. Neither self-hosting nor a provider contract automatically resolves rights in training data, confidential inputs or generated outputs.

Make switching measurable

  • Run the same representative tasks on both options, with the same acceptance criteria. Record total cost, latency, failure handling and staff time—not just benchmark scores.
  • Estimate migration work: data export, prompts, integrations, retraining and evaluation. The cheaper initial option may not have the cheaper lifetime cost.

Further reading

General information only; not investment advice, an offer or a recommendation. Any participation is subject to eligibility, documentation and applicable permissions. Capital and returns are not guaranteed; some or all capital may be lost.