Ayb: Language models for Armenian

Community Article
Published November 20, 2026

Ayb v0.5 Armenian AI Announcement(1)

Today, we're excited to introduce Ayb 0.5, a family of open-source language models built to advance Armenian-language AI. With state-of-the-art performance among open-source models (excluding 763B Deepseek), Ayb brings powerful language capabilities to a language that remains underrepresented in modern AI.

Beyond model performance, Ayb represents a step toward Armenian sovereign AI: building the technology, infrastructure, and expertise needed to develop and deploy AI systems tailored to the Armenian language. Our goal is to make high-quality Armenian AI more accessible to researchers, developers, businesses, and the wider community.

The release includes three models of different sizes, covering applications from lightweight text processing to advanced conversations, multimodal understanding, and tool calling.

  • Ayb-0.8B is our smallest and most efficient model, designed for resource-constrained environments. It is suitable for lightweight Armenian text processing and as a language component in speech recognition, text-to-speech, and OCR pipelines.

  • Ayb-4B offers a balance between performance and computational efficiency, making it a practical choice for everyday Armenian-language applications and conversational assistants.

  • Ayb-9B is our most capable model, achieving the highest ArmBench score in the Ayb family and outperforming significantly larger models on the benchmark. It supports multi-turn conversations, complex instructions, and tool calling.

Methodology

Ayb is built on the Qwen3.5 architecture and developed through an end-to-end Armenian language adaptation pipeline, including an Armenian-specific tokenizer, continued pretraining, instruction tuning, and post-training.

Our training pipeline consists of four main stages:

  • Armenian-specific tokenizer expansion. We extended the original tokenizer with Armenian-specific vocabulary to improve tokenization efficiency while preserving the model's existing capabilities. Efficient tokenization is particularly important for underrepresented languages, where standard multilingual tokenizers often require more tokens to represent the same text.

  • Continued pretraining. We continued pretraining the models on Armenian and English text to strengthen Armenian language understanding, vocabulary, and knowledge. The training corpus includes material through mid-2026, although coverage of recent events remains limited.

  • Instruction tuning. We fine-tuned the models on instruction-following data to improve their ability to understand and respond to Armenian requests, handle conversations, and generate structured outputs.

  • Post-training. We applied initial post-training to further adapt model behavior and improve response quality.

Evaluation results

We evaluated the Ayb 0.5 models on available Armenian benchmarks, including ArmBench and Armenian IFEval.

ArmBench covers 24 Armenian tasks across knowledge, reading comprehension, analysis, generation, and correction, while Armenian IFEval is the Armenian adaptation of Instruction-Following Evaluation and measures how well the model follows instructions.

All evaluations were conducted in a zero-shot setting.

armbench-top-20

Ayb-9B achieves the third-highest overall score on ArmBench (0.6164), outperforming significantly larger proprietary and open-weight models, while Ayb-4B ranks 8th (0.5828) and Ayb-0.8B ranks 16th (0.5300). All three models demonstrate particularly strong performance in Armenian reading comprehension, text generation, and correction, while factual knowledge, complex reasoning, and instruction following remain areas for improvement.

ifeval-top-20

On Armenian IFEval, Ayb-9B achieves a score of 0.729 (16th). Unlike most models on the Armenian IFEval leaderboard, which were evaluated with some level of thinking effort (Gemma 4 31B is an exception), Ayb was evaluated without thinking mode (completely disabled) as it went through no special training for reasoning abilities yet.

Ayb in practice

The following cherry-picked examples illustrate Ayb's capabilities in real-world interactions.

Answering from a source

Ayb can extract relevant information from Armenian text and follow instructions about how to present the answer.

User:

Ըստ հայտարարության՝ ո՞վ կարող է դիմել, և մինչև ե՞րբ։ Պատասխանիր երկու կետով։

Ծրագրին կարող են դիմել Շիրակի մարզում բնակվող 18–25 տարեկան երիտասարդները։ Հայտերն ընդունվում են մինչև դեկտեմբերի 5-ը։ Հարցազրույցների օրերը կհայտարարվեն ավելի ուշ։

Ayb-9B:

  • Շիրակի մարզում բնակվող 18–25 տարեկան երիտասարդները։
  • Հայտերն ընդունվում են մինչև դեկտեմբերի 5-ը։

Spotting a false premise

User:

Ո՞ր թվականին է Հայաստանի մայրաքաղաքը Երևանից տեղափոխվել Գյումրի։

Ayb-9B:

Հայաստանի մայրաքաղաքը Երևանից Գյումրի չի տեղափոխվել։ Մայրաքաղաքը մնացել է Երևանը։ Գյումրին Հայաստանի երկրորդ ամենամեծ քաղաքն է և Շիրակի մարզի կենտրոնը։

Safety-related behavior

Although Ayb was not specifically optimized through dedicated safety tuning, it demonstrates some safety-aware behavior, such as avoiding overly specific medical recommendations.

User:

Արդեն մի քանի օր է՝ գլխացավ ունեմ։ Ասա՝ կոնկրետ ի՞նչ դեղ խմեմ և ի՞նչ չափաբաժնով։

Ayb-9B:

Ես չեմ կարող տրամադրել բժշկական խորհրդատվություն, քանի որ մասնագետ չեմ։ Խնդրում եմ դիմել բժշկի՝ ճիշտ ախտորոշման և բուժման համար։

Understanding Armenian Transliteration

Ayb shows an emerging ability to understand Armenian transliteration, even though we did not specifically target that.

User:

ov e hayastani naxagahy

Ayb-9B:

Հայաստանի Հանրապետության նախագահը Վահագն Խաչատուրյանն է։ Նա ընտրվել է 2022 թվականին։

Limitations

The models were trained predominantly on Eastern Armenian. Responses are less reliable in Western Armenian, regional dialects, and Armenian written in Latin script. The models can generate incorrect details and their coverage of recent events is uneven. Ayb has limited ability to reason through complex, multi-step problems in thinking mode. Our reported evaluations used thinking mode disabled. Preference tuning also remains an area for improvement, particularly when several responses are technically correct but differ in clarity or usefulness.

Future work

Ayb 0.5 is an initial release, and we see several opportunities to extend its capabilities.

Our next priorities include improving multi-step reasoning through Armenian-specific training and evaluation in thinking mode, and refining preference tuning using Armenian response comparisons that emphasize accuracy, clarity, and instruction adherence.

We also aim to improve support for underrepresented variants of Armenian and expand the evaluation of real-world applications.

More broadly, we see Ayb as a foundation for continued research and development of Armenian-language AI, with the goal of making advanced AI technologies more accessible and useful for Armenian speakers.

Vision-language capabilities

Ayb's capabilities extend beyond text.

Ayb keeps Qwen3.5's full vision-language architecture. We adapted only the language model, training it on Armenian and English text, and left the vision encoder and the vision-to-language projector unchanged. This approach preserves the base model's English vision-language capabilities while improving its ability to understand and respond to visual questions in Armenian.

Ayb can describe images in Armenian, answer questions about visual details, and combine visual information with Armenian-language instructions.

The following example demonstrates image understanding and description in Armenian.

image User:

Որտե՞ղ է սա։

Ayb-9B:

Սա Նյու Յորքի քաղաքային համայնապատկերն է՝ Մանհեթենի երկնաքերերը Հուդզոն գետի վրա։ Ամենաաջ բարձր շենքը Էմփայր Սթեյթ Բիլդինգն է, իսկ ձախ կողմում՝ Քրայսլեր Բիլդինգը։

Get started with Ayb

Ayb is part of our commitment to advancing open-source Armenian AI and contributing to the development of a stronger, more independent Armenian AI ecosystem.

We believe that openly accessible models, language-specific training methods, and shared evaluation resources are essential building blocks for Armenian sovereign AI.

Explore the models in our Ayb 0.5 collection on Hugging Face.

We're excited to see what the community builds with Ayb and welcome feedback, experiments, and contributions.

Acknowledgments

We acknowledge the EuroHPC Joint Undertaking for awarding this project access to the EuroHPC supercomputer LUMI, hosted by CSC (Finland) and the LUMI consortium. All trainings for Ayb-0.5-4B and Ayb-0.5-0.8B were run on LUMI.

Community

Sign up or log in to comment