We develop the intelligence itself.
We work on the models themselves: architectures, weights, training and evaluation. Then we specialise them for fields where mistakes are expensive.
AI and models
The work runs from fine-tuning and specialising open models to pre-training base models of our own, and to architectures of our own design. Fine-tuning has begun; pre-training follows.
Size is a decision, not a measure of ambition. We train at the size a problem needs. A small model trained well for one field can outperform a general model many times larger.
Models are released with open weights by default, each with its evaluation and an account of how it was trained. Exceptions are made where a release would be unsafe, chiefly in defence and security.
Directions
- Base and foundation models
- Specialised models
- Original architectures
- Training and evaluation methods
Physical AI and autonomous systems
Intelligence that acts has to work in places it was not trained for.
Our Physical AI work starts in simulation: perception and control are developed there, then carried into autonomous systems, machines that operate with limited communication, energy and compute.
Weightless Intelligence works here directly. Its research into adaptation without large-scale training is aimed at physical systems first.

Cybersecurity and security intelligence
Security work is mostly reading: code, configurations, logs and traffic, in volumes no team can keep up with.
We are building specialised models for that reading, for analysis and for defensive work, and evaluate them on the tasks security teams actually perform.

Application
Defence
Synora develops AI for defence.
Perception, autonomy and decision support for systems that operate where communication is contested and energy and compute are scarce.
Defence AI for Europe should be developed and trained in Europe, under European control. That is how we build it.


Application
Enterprise
Organisations need models that know their field, not general assistants.
We build specialised intelligence for the work of organisations: systems that follow the changing information in a domain, find what matters in it and support the decision.
Each model is built for the field it serves, trained on what that field actually requires.