Kherad Afrouz

Building Persian AI, for Persian speakers and beyond.

GitHub · Hugging Face · Models · نسخه فارسی این صفحه

About Us

Kherad Afrouz (خرد افروز) is an independent, Persian-language AI initiative focused on researching, developing, and sharing artificial intelligence models and tools designed for the Persian language and its speakers.

Our long-term vision is to help reduce Iran's reliance on foreign AI technologies by building local expertise, developing specialized Persian AI models, improving access to open AI technology, and contributing to a more independent and sustainable AI ecosystem.

We believe meaningful progress does not require starting with massive models or expensive infrastructure. Small, efficient, specialized models, high-quality datasets, transparent evaluation, and reproducible research can form the foundation for more capable systems in the future.

We are starting small, learning through experimentation, and building toward a larger vision.

Our Mission

• Persian-first AI: Develop models and tools that address the linguistic characteristics and real-world needs of Persian speakers.
• Open research: Share research, code, documentation, and model artifacts whenever licensing and other constraints permit.
• Efficient models: Explore small, resource-efficient architectures that can be trained, evaluated, and deployed with limited computational resources.
• Specialized intelligence: Focus on practical tasks such as semantic search, text understanding, classification, translation, summarization, and other language technologies.
• Reproducibility and transparency: Document datasets, training methods, evaluation results, limitations, and model licenses.
• Long-term independence: Build the technical knowledge and infrastructure needed to develop increasingly capable AI systems.

Model Families

We organize our models into families, each with its own identity, purpose, and development roadmap.

Zahhak (ضحاک)

Model family archive: Hugging Face — Zahhak

Zahhak is a model-family name inspired by Persian mythology. It serves as an identity for models developed under this family.

Individual Zahhak models may target different tasks, architectures, sizes, or deployment environments. Each release should include its intended use, technical details, evaluation results, limitations, and applicable license.

The capabilities and availability of individual models depend on their respective releases.

Future Model Families

We intend to introduce additional model families as our research and engineering capabilities grow.

Model Archive

Explore our models and releases:

• All models: Hugging Face Organization
• Model repositories: Browse repositories
• Source code and tools: GitHub Organization

Each model repository should provide its own model card, usage instructions, technical specifications, evaluation results, and license information.

Our Approach

We prioritize measurable progress over model size alone.

Our development process may include:

1. Identifying a concrete Persian-language problem.
2. Reviewing existing research, models, and datasets.
3. Establishing reproducible baselines.
4. Developing or adapting suitable architectures.
5. Training and evaluating models within available compute budgets.
6. Publishing results, limitations, and reusable resources.

We do not consider a model successful merely because it has been trained or released. Its usefulness must be demonstrated through appropriate evaluation and real-world testing.

Open Source and Responsible Development

We aim to make our work accessible to researchers, developers, students, and the wider Persian-speaking community.

Availability and usage rights may differ between repositories. Please consult the license and documentation of each individual model, dataset, and project before using or redistributing it.

We value responsible data collection, respect for intellectual property, privacy protection, and honest reporting of model capabilities.

Get Involved

Kherad Afrouz is an evolving initiative, and there is room to contribute.

You can help by:

• Contributing to code and infrastructure.
• Collecting, curating, or evaluating properly licensed Persian-language datasets.
• Testing models and reporting reproducible results.
• Improving documentation and Persian-language resources.
• Sharing research ideas and identifying practical problems worth solving.

Our goal is not simply to build models, but to help build the knowledge, tools, and community needed for a stronger Persian AI ecosystem.

Repository links, model archives, and family descriptions should be updated as new projects are released.