Models and Datasets

Open source · Hugging Face

Models and Datasets

Open AI for the Macedonian language

VEZILKA develops open AI models and datasets for the Macedonian language and publishes them on Hugging Face, free for researchers, companies, public institutions and students to download, test and build on.

Browse all models ↗Browse all datasets ↗

from transformers import pipeline asr = pipeline( "automatic-speech-recognition", model="vezilka-ai/vezilka-moonshine-base-mk",)asr("lecture.wav")# → Macedonian text

Language

Macedonian

Where

Hugging Face

Areas

Text and speech

Access

Free to download

Language models

Macedonian large language models

Models trained to understand and generate text in Macedonian.

01

VezilkaLLM

Base model

A 4-billion-parameter language model trained for Macedonian, built on Gemma 3 4B. The first model in the VEZILKA LLM series.

View on Hugging Face ↗

02

VezilkaLLM-Instruct

Instruction-tuned

The instruction-following version of VezilkaLLM, for answering questions and following tasks in Macedonian.

View on Hugging Face ↗

03

Macedonian Gemma 4 E4B

Instruction-tuned

Gemma 4 E4B adapted to Macedonian, available as a pretrained and an instruction-tuned model.

View on Hugging Face ↗

04

Llama 3.2 1B · Macedonian

Compact

A small model with continued pretraining on Macedonian text, suited to experiments on modest hardware.

View on Hugging Face ↗

Speech

Speech recognition and synthesis

Models that turn Macedonian speech into text, and text into speech.

05

Whisper Large V3 Turbo · Macedonian

Speech to text

Whisper adapted to Macedonian and evaluated on real university lectures.

View on Hugging Face ↗

06

Moonshine Base · Macedonian

Speech to text

A lightweight model (61.5M parameters) for fast Macedonian speech recognition on everyday devices.

View on Hugging Face ↗

07

OmniVoice · Macedonian

Text to speech

OmniVoice fine-tuned to generate natural-sounding Macedonian speech from text.

View on Hugging Face ↗

08

Moonshine Tiny · Macedonian

Speech to text

A tiny model (27M parameters) for Macedonian speech recognition on phones and low-power devices.

View on Hugging Face ↗

Datasets

Open datasets

Data we collect and publish so anyone can train and evaluate AI for Macedonian.

09

Speech · Transcribed

Macedonian Speech Collection

Macedonian speech with transcripts for training and testing speech recognition, collected from several sources:

University lecturesParliamentary sessionsDialect fieldworkDonated read speechLecture benchmark

View on Hugging Face ↗

These are selected highlights. The full collection on Hugging Face is updated as new models are released. Each model page lists its licence, how it was trained and how to use it.

Work with us

Want to use these models in your organisation?

VEZILKA helps SMEs, startups, public institutions and researchers test and adopt AI, from a first idea to a working pilot.

Contact usOur services