AI & Compute
Open weights, closed models, and what the distinction means
Releasing model weights is not the same as open source, and the practical consequences differ from the licensing argument.
Model releases are described as open with a looseness that obscures several genuinely different things.
The gradations
Fully closed. Available only through an API. No weights, no architecture details beyond a paper, no training data information.
Open weights. The trained parameters are downloadable. Architecture is inspectable because the weights imply it.
Training data, training code and the recipe are generally not released.
Open weights with restrictions. The most common category presented as open.
Licences may restrict commercial use above a scale threshold, prohibit certain applications, or require attribution and acceptable-use compliance.
Several widely used licences of this kind do not meet the standard definition of open source, which is why the industry increasingly says open weights rather than open source.
Genuinely open source. Weights, code, training data details and an OSI-approved licence. Rare, and a small number of research efforts have done it.
Why the distinction matters practically
Not primarily an ideological question.
Reproducibility. Without training data and code, results cannot be reproduced or audited. Claims about what a model was trained on cannot be verified.
Legal certainty. A restrictive licence creates compliance risk for a business building on it, particularly where terms may change for future versions.
Fine-tuning and adaptation. Open weights allow domain adaptation, which is the main practical reason organisations want them.
Deployment control. Running weights on your own hardware means data does not leave your environment — decisive for regulated industries and for anyone with confidentiality obligations.
Longevity. An API can be deprecated. Downloaded weights cannot be taken away.
That last point is underrated. Applications built on a hosted model have been broken by model retirement or behaviour change more than once.
The safety argument, both ways
The genuine disagreement.
Against open release: once weights are public they cannot be recalled. Safety mitigations trained into a model can be removed by fine-tuning with modest effort, which has been demonstrated. Misuse becomes unmonitorable.
For open release: external researchers can audit for bias, vulnerabilities and failure modes that the developer will not find or will not publicise. Concentration of capability in a few firms is itself a risk. And the marginal uplift to a determined bad actor from an open model, relative to a closed one that can be jailbroken or to existing resources, may be small.
Empirical work on that last question exists and is not conclusive. The honest position is that the answer depends on capability level and probably changes as capability rises.
The economics
Why a company would release weights at all.
Commoditising a complement. If your business is cloud infrastructure, cheap capable models increase demand for compute.
Ecosystem and talent. Open weights attract developers, research attention and hiring interest.
Standard setting. A widely adopted open model becomes the default that tooling is built around.
Competitive positioning. Undercutting a leader's proprietary advantage by giving away something comparable.
None of these are charitable motives and all are legitimate. The result has been a genuinely competitive open ecosystem, which was not obviously going to happen.
What open weights have enabled
A substantial body of technique that would otherwise not exist.
Efficient fine-tuning methods that adapt a model with a small fraction of the parameters. Quantisation research. Inference optimisations. Interpretability work requiring access to internals. Domain-specific models in medicine, law and low-resource languages.
Almost all of it developed on openly released weights, because it requires access the API does not provide.
Derivatives and the licence chain
A practical wrinkle for anyone building on open weights.
Fine-tuning a model produces a derivative, and most licences extend their terms to it — including acceptable-use restrictions and, in some cases, naming requirements.
Which means a model fine-tuned on a restricted base carries those restrictions forward, and a product built on it inherits them.
Where several models have been merged or distilled from one another, the licence chain can become genuinely difficult to establish, and the provenance is frequently undocumented.
This matters commercially rather than philosophically, and it is worth resolving before a product depends on it.
The practical guidance
If you need data to stay in your environment, or need the model to persist unchanged, open weights are the only option.
If you need the highest capability available and can accept an external dependency, hosted frontier models generally lead.
Read the licence before building a business on it, specifically the commercial-use and acceptable-use clauses.
And distinguish, when reading announcements, between weights released and everything released. The gap between them is where most of the reproducibility problem lives.
