DATAcenter AI
The Latvian Language in the Era of LLMs: Test Results for EuroLLM-9B and TildeOpen-30B
The Latvian language is increasingly entering the world of artificial intelligence. Last month, AIdatacenterengineers.com team managed to conduct an interesting experiment-comparing the ability of two, rather small, Large Language Models (LLMs) to speak, think, and respond in Latvian.
Two models were compared in the test:
TildeOpen30B – A local development based on an extensive Latvian language dataset;
EuroLLM9B – A European-level model with an exceptionally broad multilingual knowledge base.
How the Test Was Conducted
The experiment utilized:
Python and a Gradio interface to ensure simple interaction with both models;
Several identical queries in Latvian to evaluate:
Response quality and accuracy,
Language fluency and naturalness,
The ability to follow instructions.
Results
The analysis revealed significant differences in the Latvian language performance of both models:
EuroLLM9B impressed with a very high level of Latvian language comprehension and usage. It is capable of answering both informational and conversational questions accurately without losing grammatical quality.
TildeOpen30B, on the other hand, demonstrated excellent knowledge of Latvian literature and language nuances. However, it was observed that the model lacks an instruction-following block-consequently, when assigned tasks or asked for structured answers, the model sometimes focuses on literary rather than practical phrasing.
What Does This Mean in Practice?
These results show that European-scale LLM solutions are already ready for high-quality use in Latvian. Meanwhile, Tilde’s model is highly valuable in academic and cultural contexts but still requires additional tuning to serve as practical AI assistants and chatbots.
Technologies Used
Python – for data and model integration
Gradio – for the interactive test interface
AIdatacenterengineers.com infrastructure – ensured model testing in a secure and powerful environment
What's to be tested
This test is comparing two rather small LLM’s without any detailed instructions given. Now, with the latest Nvidia DGX Spark availability you can run larger 70B+ models inside your local AI box – and buy how good they are in understanding small languages.
Contact us to implement your own, private, AI agentic system.
Datakom offers a range of IT services, suitable for any institution or company, based on business needs, industry specifics and the number of employees.
+371 67442800
© DATAKOM SIA. All Rights Reserved.