Efficient, open and sovereign.
Milmoh AI Lab is Milmoh's artificial intelligence research area. We study how to compress, optimise, align and deploy language models without giving up their capability, and how to do so on European infrastructure.
Scale is not the only direction
Over the past few years, progress in artificial intelligence has been measured mainly in scale: more parameters, more data, more compute. That path has produced highly capable systems. It has also concentrated their development in a few organisations and raised the cost of using them.
We believe there is another direction that matters just as much: understanding how much of that capability can be kept in models that are smaller, cheaper to run and deployable wherever each organisation decides, and aligned with the moral and ethical values of European society.
Milmoh AI Lab exists to work on that question, with algorithmic rigour, from Europe.
More capability with less compute.
Models and data under European control and alignment.
What is not measured cannot be taken as true.
Research areas
Five areas, one underlying question: how to achieve the greatest possible capability with the fewest resources and the most control.
Sovereign AI and open models
What does a European organisation need in order to run its own models?
We work with open-weight models, which can be inspected, adapted and run on in-house infrastructure or in European clouds. We study how to select, fine-tune and deploy them while keeping control of the data and complying by design with the GDPR and the EU AI Act.
Model distillation and compression
How much of a large model's knowledge can be transferred to a small one?
Distillation trains a compact model to reproduce the behaviour of a much larger one on specific tasks. We investigate how to combine it with quantisation and pruning to obtain specialised models that keep the quality that matters at a fraction of the memory, latency and cost.
Training and inference optimisation
Where is efficiency lost when training and running a model?
We analyse training engines, fine-tuning methods and production execution to identify where compute and energy are consumed, and how to reduce them without degrading the result.
Agentic AI
How are reliable agents built on open, efficient models?
Agents plan, use tools and complete multi-step tasks. We study what makes them reliable and supervisable, and to what extent they can run on small models rather than only on the largest ones.
Evaluation and reliability
How do we know whether a model works on the task that matters?
General metrics do not always predict performance in a real case. We develop rigorous evaluation methods, with particular attention to Spanish, as one of the languages of Europe, and to professional contexts, so that decisions about models are based on evidence.
Working on any of these questions, or interested in applying them?
Propose a collaboration →Our approach
Four principles guide the lab's work.
Open by default
We start from open models and tools whenever possible. What can be inspected can be understood, and what can be understood can be improved.
Efficiency as a metric
A model is judged not only by what it does, but by what it costs to do it: compute, energy, memory and response time.
Statistical rigour
We come from statistics and operations research, and we bring the same standards to artificial intelligence: clear hypotheses, reproducible measurements and cautious conclusions.
From Europe
We work within the European framework for data protection and AI regulation, and we see it as a starting point, not an obstacle. We also hold that the European values of democracy and society are non-negotiable conditions that must be present in the answers and reasoning of AI models.
We collaborate with universities, research centres and organisations that prefer to measure before deciding.
Write to usPrincipal investigators
Juan Aparicio
Full Professor of Statistics and Operations Research at Miguel Hernández University of Elche (UMH) and director of the University Institute for Research in Mathematics, Data Science and Artificial Intelligence (i-MInD). His research focuses on efficiency and productivity analysis; he is a co-developer of Efficiency Analysis Trees, a methodology that combines machine learning and efficiency measurement. In 2023 he was among the 25 most-cited Spanish researchers in Statistics and Optimisation. He currently leads a line of research on developing solutions based on large language models.
Xavier Barber
Full Professor of Statistics and Operations Research at UMH, where he has worked since 2003, and a member of i-MInD. He holds a PhD from UMH, and his research focuses on Bayesian statistics and spatio-temporal and longitudinal models, with more than a hundred scientific publications. He has served as Deputy Vice-Rector for Research at UMH, developed the VioGen 2 system for predicting gender-based violence with Spain's Ministry of the Interior, and taken part in projects funded by the European Commission.
Let's talk research
For research collaborations, academic enquiries or joint projects with the lab, write to us. We reply to every message.
Artificial intelligence doesn't have to be bigger to be better.
