[ About ]
TERNARY is an Estonian artificial intelligence and software technology company. We are interested in AI not only as a finished service, but as a technology that can be studied, measured, adapted and built for a specific task.
TERNARY came out of a wish to combine long experience in software development with practical AI work — from models and data to hardware, runtime and real applications.
[ Why we do this ]
We like to experiment.
With a new technology we are interested not only in what an existing solution already does, but also in whether the same task could be solved more simply, with fewer resources, or with a different architecture.
AI offers plenty of room for that. The size of a model is not a goal in itself for us. What matters more is how well the model does its job, how many resources it needs, and how controllably it can be used.
[ Small does not mean simple ]
A large model can solve a great many different tasks, but not every problem needs the largest model or a large cloud infrastructure.
We believe that with a well-chosen architecture, good data and enough specialisation, even a relatively small model can be very capable in practice.
Such a model can run locally, needs less memory and compute, and is easier to control and to integrate.
This is also what we study in our MATRIX technology: what is the smallest model that is good enough for a given task?
[ A model should be as large as necessary, not larger. ]
[ 25+ years of IT ]
Behind TERNARY there are more than 25 years of software development and IT experience.
Over that time the technologies have changed repeatedly — from the web and server applications to mobile, the cloud, and now artificial intelligence. The principle has stayed the same: technology has to work in a real system and solve a real problem.
In AI work that means a model alone is not yet a solution. It needs data, software, a runtime environment, interfaces, testing and often the right hardware as well.
[ Our own technology ]
TERNARY does not limit itself to using ready-made models and cloud services.
We also develop our own technology: Estonian speech synthesis HÄÄL, speech recognition KUULA, and a technical direction for compact models, MATRIX.
These projects let us work with the whole chain — from training data and model architecture to quantisation, runtime, local operation and applications.
The aim of that work is not to copy existing solutions. We want to understand where something can be done in a smaller, simpler or more controllable way.
[ Research and collaboration ]
In the next stage of MATRIX we want to involve independent testers, and to make the model available for evaluation to universities and research institutions.
The aim is to obtain external measurements, feedback and comparisons that do not rest on our own development team's judgement alone.
We are also looking for ways to work with research and development institutions where the MATRIX architecture, training methods, properties of local use or applications of specialised models could be studied through joint experiments and projects.
It matters to us that the progress of a model can be judged from reproducible tests and independent feedback, not only from internal examples.
[ Direction ]
Our long-term aim is to help shape a direction in which practical AI does not automatically mean ever larger models and more powerful hardware.
We want to develop small, well-trained and specialised models that can be used locally, and whose behaviour, cost and data flow can be controlled.
For some tasks a large model is the right choice. For others a much smaller one is enough. It is that second part that interests us most.
[ How we work ]
We experiment, measure and compare.
If an existing model fits, we use it. If it makes sense to adapt it, we adapt it. If the task justifies a small model of its own, we train one.
The same principle applies to infrastructure: