Open-Weight Models and the Limits of the Word Open
Downloadable weights are not the same as open source, and the difference has legal consequences.
"Open" is doing a great deal of work in the phrase "open model", and most of it is marketing. The distinction that matters to anyone deploying one is not philosophical but contractual, and it is visible only in the licence file that almost nobody reads before the pilot starts.
The category commonly described as open contains at least three quite different things. There are models released under recognised open-source licences such as Apache 2.0 or MIT, where use, modification and redistribution are permitted with minimal conditions. There are models released under bespoke community licences that permit broad use but attach conditions — acceptable-use policies, scale thresholds above which a separate agreement is required, restrictions on using outputs to train competing models, naming and attribution obligations. And there are models whose weights are downloadable but whose licences are, in substance, restrictive commercial agreements distributed through a public download.
Only the first is open source by the definition the term has carried for twenty-five years. The Open Source Initiative, which maintains that definition, does not recognise licences that restrict fields of use. The second and third categories are better described as open-weight: you can hold the artefact, run it where you choose, inspect its behaviour and fine-tune it, and those are real and substantial freedoms. What you cannot necessarily do is everything the word "open" implies.
There is a further gap that applies to nearly all of them. Open weights are not open training data. Very few releases disclose the corpus, its provenance or its licensing status, which means a downstream user cannot independently assess copyright exposure, contamination of evaluation benchmarks, or whether restricted material was included. Reproducibility — the property that made open source valuable to science — is largely absent. You can rerun the model; you cannot rebuild it.
Why this matters practically comes down to four questions a deploying organisation should answer before the model is in a product. Does the licence restrict the field of use in a way that touches the intended application, including sectors the company might enter later. Is there a user or revenue threshold above which separate terms apply, and is the business plan comfortably below it. Does the licence constrain what may be done with outputs, particularly using them to train or improve another model. And what indemnity, if any, exists for intellectual-property claims arising from generated output — for hosted commercial models this is increasingly offered; for downloaded weights it generally is not, and the exposure sits with the deployer.
The strategic reading is that open-weight releases are, for the companies publishing them, a commercial instrument rather than an act of charity. They commoditise a competitor's paid layer, seed an ecosystem, recruit developers and generate evaluation signal at scale. That does not make them less useful. It does mean the terms are set to serve the publisher's position, and can be revised for subsequent versions — an important point for anyone treating a current licence as a permanent property of the model family.
The limitations of this analysis are worth marking. Licence terms change between releases and the summary above describes categories rather than any specific agreement; the only reliable source is the licence text attached to the version being deployed. Interpretation of several common clauses has not been tested in court in most jurisdictions, so reasonable counsel disagree. And regulatory treatment of open-weight releases is in flux, with obligations in some jurisdictions turning partly on whether a model is freely available, which creates the odd situation where the licence affects the compliance burden.
The useful discipline is simply to stop using the word. Ask instead: what may we do with this, at what scale, in which sectors, with whose indemnity, under terms that may change at the next release. Those questions have answers. "Is it open" does not.
Every claim above is sourced to a document, a named person, or a record we hold. Where we could not verify a claim, we say so. Read our standards and corrections policy →
