Overview
Today we're releasing Ayb v0.5, our open large language model for Armenian and our first step towards sovereign Armenian AI. Its largest version, Ayb 9B, is the state-of-the-art open-source model for Armenian, and on Armenian tasks it competes with, and often beats, the largest frontier models from Google, OpenAI and Anthropic.
Ayb was trained end to end, from pretraining to reinforcement learning, on top of a tokenizer we built for Armenian. That makes it much cheaper and faster to run on Armenian text than closed models such as Claude. It also comes with early vision capabilities, and in three sizes: 0.8B, 4B and 9B.
Այբ (Ayb) is the first letter of the Armenian alphabet, created by Mesrop Mashtots in 405 AD. Written Armenian begins with it, and so does our Armenian AI stack.
Sovereign Armenian AI
Until now, using a top-tier language model in Armenian meant sending your data to a foreign API, paying per token for a language those models were never optimized for, and depending on decisions made somewhere else. For governments, banks, hospitals and schools that is often not an option at all.
Ayb changes that. The weights are open, so the model can run on your own infrastructure, in Armenia, be audited, and be fine-tuned on your own data, without giving up quality. Ayb v0.5 is the first release on that path. Together with OhanASR and OhanTTS, it covers listening, reasoning and speaking in Armenian on open models.
Three sizes
All three sizes are in one Hugging Face collection, so you can pick the balance of quality and cost that fits your use case:
- Ayb 9B is the flagship and the state-of-the-art open-source Armenian model. It goes head to head with Gemini, GPT and Claude on Armenian tasks.
- Ayb 4B is the middle ground: strong Armenian quality for a fraction of the compute.
- Ayb 0.8B is small enough for laptops, edge devices and high-volume pipelines.
Trained end to end
Ayb is not a thin fine-tune on top of someone else's chatbot. We trained it through the full pipeline:
- Pretraining on Armenian text, so the model learns the language, its grammar and its knowledge at the foundation.
- Supervised fine-tuning (SFT) to follow instructions and hold natural conversations in Armenian.
- Preference optimization (DPO) to align its answers with what Armenian speakers actually prefer.
- Reinforcement learning (GRPO) to sharpen reasoning and accuracy on verifiable tasks.
An Armenian tokenizer
Most LLMs tokenize Armenian badly. Their vocabularies were built mostly for English, so Armenian words get split into many small pieces. Every extra token costs money, adds latency and uses up context.
We built Ayb's tokenizer to be efficient for Armenian. The same Armenian text takes far fewer tokens, so Ayb is significantly cheaper and faster to run than closed models such as Claude, and fits more Armenian into the same context window.
Early vision
Ayb also understands images. Through a lightweight fine-tuning stage, it gained early vision capabilities: reading Armenian text in photos and screenshots, pulling information out of forms and scanned documents, and answering questions about images. This is an early capability that we'll keep strengthening in future releases.
Evaluation
We evaluate Ayb on ArmBench-LLM, the open Armenian LLM benchmark we built and maintain. Every number is reproducible and directly comparable with open and closed models on the public leaderboard.
Get the models
All three sizes are available on Hugging Face in one collection. For commercial licensing, fine-tuning or on-prem deployment, get in touch.
What's next
Ayb v0.5 is the first step. Next we're working on stronger reasoning, longer context, more capable vision, and better handling of Western Armenian and dialects. We'll share the research behind each release on this blog.



