We train
MINA's weights ourselves.
We are developing MINA as a clinical AI model for the Turkish healthcare system. What a language model knows sits in its weights. A company that trains the weights itself decides what the model learns and where the model runs.
We train the new version in two stages
We train the new version of MINA in two stages. In the first stage, the model reads clinical guidelines, drug information and the literature, and builds clinical knowledge into its weights. In the second stage, we teach the model to reason over a case, to state its decision with its rationale and basis, and to say so when it is unsure.
The model reads frequently changing information, such as SUT articles, tariffs and codes, from current sources through tools. When this information changes, the model does not need to be retrained. Institution-specific rules do not go into the weights of MINA's shared model; MINA reads these rules only through retrieval-augmented generation (RAG).
Our decision approach comes from the field
Over five years of development, the decision approach we teach the model has come from the field. We are teaching MINA to look at the same file through the eyes of the physician, the hospital, the insurer and the inspector. You can find the results of a measurement that looks at the same file from different roles in The same file, four pairs of eyes.
The team behind the work
Data preparation, tracing the source of every training example, training on GPU infrastructure and measuring every release each require their own expertise. This expertise builds up only through years of experience. For MINA, we trace the source of every training example and measure every release.
Our 17-person team, from AI researchers to clinical domain experts, works on MINA. The Physician Ethics and Advisory Board, made up of physicians from 15 specialties, approves model updates. An approved release is deployed and runs on the institution's own servers or in an isolated cloud environment.
You can read how MINA was built on the MINA page. If you would like to see it in your own institution, you can request a POC.
This article describes the training approach we adopted for MINA's new version in October 2026. The evaluation results published on this site refer to MINA v3.7.
SSS
How is MINA's new version trained?
The new version is trained in two stages. In the first stage, the model reads clinical guidelines, drug information and the literature, and builds clinical knowledge into its weights. In the second stage, it learns to reason over a case, to state its decision with its rationale and basis, and to say so when it is unsure.
Where does MINA get frequently changing information such as SUT and tariffs?
MINA reads frequently changing information from current sources through tools. Institution-specific rules do not go into the weights of MINA's shared model; MINA reads these rules only through retrieval-augmented generation (RAG).
Who approves model updates?
The Physician Ethics and Advisory Board, made up of physicians from 15 specialties, approves model updates. A release measured with evaluation protocols is deployed in the institution's own environment after the board approves it.