About ufak AI
ufak AI is an independent research lab founded in Türkiye in 2026. “ufak” is Turkish for small. We train small language models from scratch and publish how they were trained and measured.
Turkish comes first. For our first model, ufakzeka-1, the tokenizer was trained on Turkish text, most of the pretraining data was Turkish, and the model was evaluated on Turkish tests.
The weights of ufakzeka-1 are released under Apache-2.0. The GitHub repository holds the data preparation, training and evaluation code, the list of data sources and the spend ledger.
The second project is ufakzeka-karar, a Turkish decision model built on ufakzeka-1-base, together with HakemBench, the Turkish benchmark it is measured on. The model weights, the whole open benchmark of 2,346 items, the scoring code and the leaderboard are released together.
The third project is ufakzeka-kulak, a small Turkish speech recognition model trained from scratch on speech recordings whose licences allow commercial use. Its weights are released under Apache-2.0 for five runtimes, and its demo runs in the browser.
If you can offer GPU access or want to support open research, write to us.