Open-source AI is not free: the costs and responsibilities to plan
Short answer: Open source, open weights and source-available grant different rights over code, weights and redistribution. A zero licence fee does not remove hardware, security or ownership work.
Make the terms precise
Read the licences for code, model weights and datasets separately. Record source, version, model card, notices, data rights and commercial restrictions in the delivery inventory.
Costs to include
| Item | Question to answer |
|---|---|
| runtime/hardware | Who tests compatibility and owns the accelerator? |
| security/SBOM | Who tracks CVEs and patches upstream changes? |
| evaluation | Can a replacement use the same API contract and acceptance method? |
| handover/exit | Are the runbook, inventory, rollback and receiving owner ready? |
Use the Open-source AI pillar to compare choices and the Security policy to bound the operating work.
Portability needs evidence
Keep the application contract separate from provider-specific runtime. Test a replacement with the same workload. A downloadable model is not proof that it runs on every hardware target.
Limitations
Upstream projects can stop support or change a licence, a runtime may not support the same precision, and training-data rights remain with their owners. Handover must list artefacts and rights item by item.
Read next
Local AI or Cloud AI: an organisation decision tablePrivate RAG: a document-search architecture that keeps data in the boundaryA Production Readiness Review for an AI systemContinue with the decision context
Want to talk this problem through with our team?
A free 30-minute conversation, no commitment, no jargon you cannot follow.
Talk to us for 30 minutes first