Ukrainians have created a new language model for the Ukrainian language, MamayLM, which surpasses all similar models in size for both English and Ukrainian, and is also capable of outperforming models up to 10 times larger (for example, on external assessment tasks).
Ukrainians have created a new language model for the Ukrainian language, MamayLM, which surpasses all similar models in size for both English and Ukrainian, and is also capable of outperforming models up to 10 times larger (for example, on external assessment tasks).
As the development team told dev.ua, the team is part of the INSAIT Institute in Bulgaria and ETH Zurich in Switzerland.
«MamayLM is cost-effective and can run on a single GPU. It is ideally suited for enabling fast, cost-effective autonomy and AI implementation in various sectors of Ukraine, including education, law, healthcare, and the public sector (e.g., by specializing the model for different use cases). It is also attractive in scenarios where data confidentiality is important, as it can be run on a local machine,» the inventors said.
The team has already held talks on planning cooperation with the Ministry of Digital Transformation. «We are in touch with the AI Center of Excellence team, which is currently focused on launching the development of a large-scale Ukrainian language model. Our cooperation will focus on exchanging experience and joint testing to develop and assist in the implementation of AI in the country. We also include in the copy the Embassy of Ukraine in Bulgaria, which was informed in advance about this important progress for Ukraine,» the developers reported.
MamayLM is freely available on HuggingFace.
MamayLM is built on the Google Gemma 2 9B model, which INSAIT previously used to develop the BgGPT 2.0 series of models. It is adapted to the specifics of the Ukrainian language, is an expert in native language and cultural nuances. The model is also a powerful foundation for creating applications and integrating them into government agencies — especially in environments where data confidentiality is critical (since such a small model can run locally) — as well as for personal use at low cost.
MamayLM demonstrates the highest score on EIA (External Independent Assessment) exams among models of similar size, outperforming even significantly larger models, including Gemma2 27B, Llama 3.1 70B, and Qwen 2.5 72B.
