Introducing Mijatovic-Adriaticum-1B

Today, Mijatovic is introducing Mijatovic-Adriaticum-1B, an independent 1B-parameter language model focused on Serbian, Croatian, Bosnian, and the wider South Slavic language space.

Adriaticum was trained from scratch and developed as part of the wider Mijatovic project. The goal is not to compete with frontier-scale models, but to build something useful for our language space — with open weights, transparent provenance, and room to improve.

Built independently

Adriaticum was built independently by one person, with feedback from a small group of testers.

The model went through several stages of pretraining, adaptation, and supervised fine-tuning using a mixture of web text, parliamentary material, translation data, instruction datasets, synthetic alignment data, and first-party alignment data.

Production Mijatovic user conversations were not used to train this release.

Open weights

Mijatovic-Adriaticum-1B is released with open weights under the Apache License 2.0.

The release includes:

• approximately 1.03 billion parameters
• a 2,048-token context window
• FP16 SafeTensors weights
• a Hugging Face-compatible chat template
• custom Transformers model code
• support centered on Serbian, Croatian, and Bosnian

Third-party training-source attribution and notices are included with the model release.

Try Adriaticum

You can try Adriaticum directly through Mijatovic Chat:

https://chat.mijatovic.io

The model and its weights are available on Hugging Face:

https://huggingface.co/mijatovic/Adriaticum-1

Adriaticum is a first public model release, not a finished destination. Feedback is welcome — especially examples of where the model works well, where it struggles, and what you would like to see improved.

Designed for everyone.

— Mijatovic