Azerbaijani at the core
A product-level approach to the expressions, terminology and local context of our language.
Azerbaijan's national artificial intelligence model
Khazri is a large language model family built for local language, cultural context and real business needs. Our goal is to strengthen Azerbaijan's role in the AI ecosystem and reduce dependence on foreign infrastructure.
Current flagship · Khazri 2 · SYNAPSE architectureExplore models
AI infrastructure built for Azerbaijan
Khazri is more than a text-generation model. It is an initiative to create a local technology base for products, services and research that work with knowledge in Azerbaijani.
Independent infrastructure
A model family designed to reduce the reliance of local applications on foreign models and services.
Modern architecture
The Khazri 2 family uses modern attention and activation methods together with the SYNAPSE architecture.
Stages of Khazri model development
From the flagship Khazri 2 to the open-source first-generation Khazri 1, each model serves a different need.
Flagship · open weights
Khazri 2
250Mparameters
The current flagship of the Khazri family, published on Hugging Face for more advanced Azerbaijani understanding and generation tasks.
Read moreCompact · open weights
Khazri 2 Mini
100Mparameters
The compact Khazri 2-family model, published on Hugging Face for applications that require less compute and lower latency.
Read moreFirst generation · open source
Khazri 1
47.7Mparameters
The first and most limited GPT-2-based model in the family, published on Hugging Face for open Azerbaijani research and experimentation.
Read moreBest use-case examples
Khazri can be adapted to workflows that require Azerbaijani-language knowledge and integrated with an organisation's own data.
Citizen services
Digital assistants that understand Azerbaijani queries and explain complex rules in plain language.
Enterprise knowledge assistant
Search, summarisation and question answering across internal documents, procedures and databases.
Document processing
Extraction, classification and structured summaries for financial, legal and operational documents.
Education and culture
Azerbaijani explanations, personalised learning materials and support for cultural content.
KHАZRI 2 PREVIEW
Strong results in early testing
Khazri 2's preview configuration was tested on selected tasks including context extraction, arithmetic, word problems and abstention.
| BenchmarkHigher is better ↗ | ♛ Best overallKhazri 2(preview) | GGemma 3270M · Google | QQwen 2.50.5B · Alibaba Cloud | PPythia160M · EleutherAI |
|---|---|---|---|---|
| Context extraction | 100%♛ | 100% | 89% | 22% |
| Mixed speed / proxy | 62%♛ | 36% | 45% | 10% |
| Arithmetic | 99%♛ | 0% | 14% | 0% |
| Word problems | 99%♛ | 0% | 28% | 2% |
| Abstention | 97.4%♛ | 18% | 46% | 1% |
This visual shows internal preview measurements supplied by the Khazri team. The results are not an independent certification, apply to a selected test configuration and may change before release.
QUICK ANSWERS
Frequently asked questions about Khazri
Core answers about the model family, access and integration.
What is Khazri?
Khazri is a large language model family built for Azerbaijani and local context. Its purpose is to strengthen Azerbaijan's AI ecosystem and reduce dependence on foreign infrastructure.
Is Khazri Azerbaijan's first LLM?
The Khazri team presents the project as Azerbaijan's first local large language model family. Khazri 1 is its first generation; Khazri 2 and Khazri 2 Mini are the newer models published on Hugging Face.
Which model is the current flagship?
The current flagship is the 250M-parameter Khazri 2. It uses modern attention and activation methods together with the SYNAPSE architecture.
What is Khazri 2 Mini designed for?
The 100M-parameter Khazri 2 Mini is published on Hugging Face for applications that need less compute, lower latency and more compact deployment.
Is Khazri 1 open source?
Yes. The GPT-2-based Khazri 1 is available on Hugging Face under GPL-3.0. Its capabilities are substantially more limited than those of the new Khazri 2 family.
KHАZRI BLOG
Notes on local artificial intelligence
Updates on the model family, its architecture direction and Azerbaijan's AI infrastructure.
Architecture
SYNAPSE: the architecture direction of the Khazri 2 family
An overview of the role modern attention and activation approaches play in the Khazri 2 family.
4 min readRead more
Ecosystem
Why does Azerbaijan need local AI infrastructure?
The role of local large language models in language coverage, data sovereignty and technological resilience.
5 min readRead more
Models
Khazri 2, Khazri 2 Mini and Khazri 1: what is the difference?
A concise comparison of the three models by generation, size, purpose and availability.
4 min readRead more
WORK WITH KHАZRI
Bring Azerbaijani AI into your product
Contact our team on WhatsApp to discuss a Khazri 2 or Khazri 2 Mini integration.
Contact us