Most "Ukrainian" AI products are wrappers around the OpenAI API with a Ukrainian interface. Lapathoniia Chat claims to be different: Ukrainian models, Ukrainian servers, data does not cross borders. How important this is depends on who uses the tool and for what purpose.
What's inside
The platform is built on two large language models of proprietary development. Lapa is a 12B Instruct model with vision support (image recognition). Mamay is a 12B Instruct model based on Gemma-3 architecture, optimized for the Ukrainian language. Both support context windows up to 32,000 tokens — enough for analyzing long documents within a single query.
The infrastructure foundation is De Novo cloud with NVIDIA H100 GPUs. The developers also provide an OpenAI-compatible API, which means Lapathoniia can be integrated into existing products without rewriting code — just by changing the endpoint.
Why "where data is stored" is not marketing
For a private user, the question of server location sounds abstract. For a government agency, medical institution, or defense contractor — it is a legal requirement and a security issue. These sectors are the obvious target audience for Lapathoniia, even if the developers publicly speak about "demonstrating capabilities."
"The emergence of MamayLM is an incentive for startups and independent developers to create projects with high privacy without dependence on cloud services"
— MamayLM review, itc.ua
Closed beta: what does it mean in practice
Currently, the service operates in closed beta mode — developers are gathering feedback and testing load before public launch. At the same time, the platform already provides API keys, budgets, and limits for teams, suggesting a commercial model rather than just a research project.
The real challenge is not technical
12B-parameter models are mid-range by global standards of 2025–2026. GPT-4o, Claude 3.5, and Gemini 1.5 Pro operate on significantly larger architectures and have years of RLHF training on billions of queries. Lapathoniia does not compete with them on generation quality — and does not appear to set such a goal.
- Advantage: sovereign data storage, Ukrainian language optimization, OpenAI-compatible API for easy integration
- Limitations: closed beta, smaller model scale compared to global leaders, unknown quality on specialized tasks
- Unknown: pricing model and access conditions after beta exit
If Lapathoniia opens access and publishes benchmarks on Ukrainian language tasks — it will become clear whether this is infrastructure sovereignty with real quality, or a compromise that only those whose laws prohibit using foreign clouds will agree to.
