Decision
Do Machine Learning Models Qualify for the Cyprus IP Box?
Do Machine Learning Models Qualify for the Cyprus IP Box?: short answer
Last reviewed
A model can qualify where it rests on a legally protected asset, most often copyright in the training and inference code, and where the company funded the development. Trained weights alone sit on weaker ground than the code that produces them, so the claim is usually built around the system rather than the model file.
| Most defensible asset | Copyright in training, inference and supporting code |
|---|---|
| Less settled | Trained parameters considered in isolation |
| Registration required | No. Copyright arises on creation |
| Effect of a third-party base model | Narrows ownership to the company's own contribution, governed by the base licence |
| Deciding variable for benefit | The nexus fraction, driven by who funded development |
| Excluded | Brand, domain value and other marketing intangibles |
AI founders assume the model is the asset. For qualification purposes the defensible asset is normally the codebase around it, which changes what has to be documented.
What the asset actually is
The regime requires a legally protected intangible asset arising from research and development. For a machine learning business, three things get conflated and they are not equivalent.
The code. Training pipelines, data processing, evaluation harnesses, inference servers and the infrastructure around them attract copyright on creation. This is the most defensible asset and the one a claim is normally built on.
The weights. A trained parameter set is the output of running that code over data. Whether weights in isolation attract copyright is less settled than it is for source code, because the parameters are produced by a process rather than authored directly.
The data. Datasets may attract protection where they involve sufficient selection or arrangement, but a corpus assembled by scraping is a weaker foundation than a curated one, and the rights in the underlying material may not belong to the company at all.
The practical consequence is that the claim is framed around the copyrighted system that produces and serves the model, rather than around a model file.
Base models change what is owned
Most products are not trained from nothing. Where a company fine-tunes or builds on a third-party base model, the licence governing that base model determines what the company can claim.
Three questions decide it:
- Does the licence permit commercial use, and on what terms?
- Does it restrict derivative works or impose conditions on outputs?
- What does the company actually own, as against what it is permitted to use?
Building on a permissively licensed base does not defeat a claim. It narrows the asset to the company's own contribution: the fine-tuning code, the data pipeline, the evaluation and serving systems, and the adaptations made. That contribution can be substantial, and it is what the documentation has to isolate.
The nexus fraction still decides the outcome
Qualification is the threshold question. The size of the benefit is a separate one, and for AI companies it is usually where the value is won or lost.
Compute is the distinctive feature. Cloud spend consumed in training and experimentation is development expenditure, and where the company incurs it directly it is qualifying expenditure. Where compute is purchased through a related entity and recharged, it enters overall expenditure without entering the qualifying numerator and dilutes the fraction.
Given how large training budgets can be relative to salaries, that classification often moves the nexus fraction more than the engineering payroll does.
The evidence is ordinary for a competent team:
- repository history showing who wrote what and when
- experiment tracking with runs, hyperparameters and outcomes
- evaluation results and the decisions taken on them
- cloud billing tied to specific training and development activity
- deployment and inference logs demonstrating the system in operation
Common questions
Are model weights themselves a qualifying asset?
Weights considered in isolation sit on less settled ground than source code, because they are generated by a process rather than authored. The stronger position frames the asset as the copyrighted training and inference system that produces and serves them, with the weights as its output.
Does fine-tuning an open-weight model disqualify the claim?
No, but it narrows what is owned to the company's own contribution: the fine-tuning code, the data pipeline, the evaluation and the serving systems. The licence governing the base model has to be read first, because it determines what may be commercialised and what rights attach to derivatives.
Does cloud compute count as qualifying expenditure?
Compute consumed in training and experimentation is development expenditure, and where the company incurs it directly it qualifies. Where it is bought through a related party and recharged, it enters overall expenditure without entering the qualifying numerator and dilutes the nexus fraction.
What if the training data was scraped?
That is a rights question before it is a tax question. A corpus assembled from third-party material may carry restrictions that affect what the company can claim to own, and it is a weaker foundation for an asset than a curated or licensed dataset.
Technical definition
A qualifying intangible asset must be legally protected and arise from research and development activity. Source code for training pipelines, inference systems and supporting infrastructure attracts copyright on creation. Trained parameters occupy less settled ground, so the qualifying asset is ordinarily framed as the copyrighted system that generates and serves them.
Practical implications
The evidence that matters is the same evidence a serious engineering organisation already produces: repository history, experiment tracking, evaluation records and deployment logs. Where a model is fine-tuned from a third-party base, the licence terms of that base model bear directly on what the company can claim to own.
Common misconceptions
The most common error is treating a set of weights as the asset. The second is assuming that using an open-weight base model is fatal to a claim. It is not, but it narrows what is owned to the company's own contribution, and the licence governing the base model has to be read before anything is asserted.