Synora is building its own models.
This week we begin fine-tuning and specialising open models, working directly with weights, data and training. Pre-training base models of our own follows, on the evidence of those experiments. The models are for fields where mistakes are expensive: cybersecurity, defence, physical systems and enterprise. They will be released with open weights by default, and the first within weeks.
This is a change in my own thinking.
For most of this year, my work went into Weightless Intelligence, on the conviction that intelligence is more than training ever-larger models. Our experiments with systems that have no pretrained weights gave me good reason to hold it. I still do.
I also concluded that I was looking at the problem from one side only. Most of today's capability lives in the current paradigm, and the best place to understand it and improve on it is from inside. So Synora's research now runs in two lines: models, inside the paradigm, and Weightless Intelligence, outside it. The two inform each other.
Size is not the point. A model trained well for one field can matter more than a general one many times its size. A breakthrough can come from a model of a billion parameters. We train at the size the problem needs.
I believe in open models. Access to capable AI should not rest with a handful of providers, and a company that asks for trust should show its work. Our models will be released with their weights, and our research with its methods. The exceptions will be few and stated, mostly in defence and security.
This is a risk, in engineering and in reputation. Work at the level of the models is expensive, slow, and public when it fails. Synora is founder-funded, with no outside equity, and has decided to take it.
We are not doing it alone. Our compute runs in Stockholm and Frankfurt on Amazon Web Services, which supports this work with cloud credits and GPU capacity. NVIDIA supports our Physical AI work, which is built on its Isaac simulation platform. I am grateful to both.
Europe needs the capacity to create foundational technology, not only to use it. That capacity is built by companies willing to take the risk.
From today, Synora publishes its work as it happens: models, experiments and papers. This statement is the first entry in that record.
Europe's AI future should be built in Europe. Bulgaria should have a place among those building it.
Krasimir PetrovFounder, Synora