Open-weight AI: inspect the release before calling it open source
Weights are only part of a release
A downloadable set of model weights does not by itself establish that a system meets the Open Source Initiative’s AI definition. The definition addresses the freedoms to use, study, modify and share, with the preferred form for modification. Check the particular release and licence rather than inferring rights from a model family name or a download button.
A release record worth keeping
| Field | Record |
|---|---|
| Identity | Repository, exact revision, model card and file checksums |
| Rights | Licence version, commercial use, redistribution and restrictions |
| Dependencies | Tokenizer, code, adapter and data information; their separate terms |
| Operation | Memory, runtime, hardware, context and network requirements |
| Governance | Update process, security response and who accepts the output |
Preserve the licence supplied with the version you actually downloaded. A later model with a similar name can have different terms. Ask a qualified adviser where rights are unclear; this article does not determine the legality of a particular deployment.
A small deployment trial
For a fictional internal draft assistant, begin with synthetic documents and a fixed set of acceptance rules. Record the exact model, prompt, runtime and whether external services are enabled. Test a request with no retrieved answer: a useful result should say the evidence is missing rather than inventing a reference. Log output quality separately from memory use and latency. This is a proposed procedure, not a benchmark we ran.
Local does not answer every privacy question
A local inference process may still call a cloud feature, download files or send application logs elsewhere. Map those paths and their settings before using sensitive records. Licence suitability, task performance and privacy behaviour are independent questions. A model can pass one and fail another; “open” does not settle all three.
Sources and scope
- OSI: Open Source AI Definition 1.0 — A definition and its requirements; not certification of every open-weight model.
Sources checked October 10, 2026. AI-assisted research and drafting. The examples, calculations and worksheets are Luminesca editorial work; they are not measured trials, first-hand reporting or an independent certification. Removed broad vendor licence claims and promises that local deployment automatically ensures privacy. Report a specific correction with its source.