Bigger doesn’t always mean better: the role of specialised models in clinical AI

In recent years an almost automatic association has taken hold between number of parameters and capability: if a model is larger, it should solve any problem better.

In artificial intelligence applied to the clinical field, that conclusion is too simple.

A medical consultation is not a single AI task. Within one workflow there can be speech recognition, identification of clinical entities, negation detection, classification, evidence checking, structuring of information and, finally, text generation.

Each problem has different characteristics.

And the available evidence suggests that, for some of those tasks, a model trained specifically to solve a particular problem can compete with much larger systems and, under certain conditions, outperform them.

Specialisation versus general capability

A study published in Nature Communications by Chen et al. in 2025 compared four large language models on twelve biomedical language processing datasets covering six types of applications.

The authors found that traditional fine-tuning approaches — including models from the BERT and BART families — outperformed large models used in zero-shot or few-shot settings on a good share of the tasks evaluated.

The advantage of large models appeared especially when the task required more open-ended reasoning, such as certain medical questions.

The work also documented an issue particularly relevant to sensitive applications: generative systems could produce seemingly convincing answers that nevertheless omitted information or introduced content not present in the original data.

This does not mean one approach is universally better than another. It means that the right architecture depends on the job it is asked to do.

Verifying a claim is not the same as writing an answer

This difference is clearly visible in biomedical verification tasks.

In a paper presented at the BioNLP Workshop at ACL 2026, Kumar studied models fine-tuned using only 1,008 training examples and compared them with large proprietary models.

On HealthVer, BioLinkBERT reached a macro-F1 of 62.4, against 53.2 and 42.4 for the two proprietary models evaluated. A 7-billion-parameter model fine-tuned with QLoRA reached 65.2.

On SciFact, BioLinkBERT scored 87.5, against 85.6 and 77.9.

The numbers should not be read as a universal ranking of models. They are results within specific datasets, configurations and tasks.

In fact, the paper itself detects a structural artefact in SciFact that may favour the metrics of the fine-tuned models. For that reason, HealthVer makes a particularly interesting comparison.

The important conclusion lies elsewhere: it is not always necessary to use the largest system available to make a bounded decision.

A quick filter before an expensive operation

Abbes et al. studied another relevant problem in 2025: determining whether the information in a document is sufficient to answer a query.

In this type of groundedness classification, lightweight encoders such as RoBERTa or NomicBERT achieved results comparable to much larger models, but with considerably lower latency.

The architecture that emerges from this kind of research is interesting: use small, fast models to filter, check or classify first, and reserve the more expensive generative models for the stages where they really add value.

Instead of asking which model can do everything, the question becomes:

which model is most appropriate for each stage?

Nor should small models be idealised

The evidence does not allow us to conclude that specialised encoders are always superior.

In 2024, Hu et al. compared BERT models with instruction-tuned models from the LLaMA family for clinical information extraction.

The LLaMA models achieved moderate improvements in entity recognition and relation extraction and showed a greater advantage when facing data not seen during training.

However, that improvement came at a cost: higher computational requirements and, in certain configurations of the study, execution up to 28 times slower.

The decision again depends on the application.

Accuracy, generalisation, latency, cost, memory consumption and auditability are all part of the same engineering problem.

What this means for a clinical AI architecture

This discussion has practical consequences.

In clinical applications, an architecture should not be designed around the model that makes the most headlines, but around a sequence of clearly evaluable tasks.

A specialised system can be especially interesting when you need to:

• extract structured information;
• detect certain entities;
• verify claims;
• classify content;
• work with low latency;
• reduce computational cost;
• or run certain processes on local infrastructure.

Larger generative models continue to play an important role when the task requires more open-ended interpretation or complex generation.

Both approaches can coexist.

The line we follow at Myukura

At Myukura we start from a similar idea: the quality of a clinical application cannot be measured solely by the name or size of the model it uses.

What also matters is the architecture around the model, the verification mechanisms, the traceability of data, the handling of errors and the professional’s involvement.

That is why we develop and evaluate both generative models and smaller, specialised systems aimed at different tasks within the clinical workflow.

We are not trying to get a single model to do everything.

Our work consists of developing an architecture in which each component serves a specific function and can be evaluated independently before being incorporated into a supervised clinical workflow.

The aim is to combine accuracy, efficiency, traceability and control, selecting for each stage the most appropriate architecture and level of specialisation.