Fanar
Arabic-centric multimodal generative AI platform and models
Fanar is a generative artificial intelligence platform focused on Arabic language and culture, developed by the Qatar Computing Research Institute. It features two large language models, Fanar Star and Fanar Prime, trained on extensive Arabic data to support text, speech, and image tasks while aligning with regional values.
- Trained on 1 trillion tokens with focus on Arabic
- Supports Gulf, Levantine, and Egyptian dialects
- Includes specialized Islamic retrieval augmented generation
- Developed by Qatar Computing Research Institute
- Offers both 7B and 9B parameter model variants
What it is
Fanar is a comprehensive artificial intelligence platform designed specifically for Arabic-speaking users. Developed by the Qatar Computing Research Institute, it integrates large language models with capabilities for speech recognition, voice synthesis, and image generation. The system aims to provide sovereign AI technology that respects local cultural and religious contexts. It supports both Modern Standard Arabic and various regional dialects, offering a tailored experience for users in the Arab world.
How it came about
The project was undertaken entirely at the Qatar Computing Research Institute, part of Hamad Bin Khalifa University. Sponsored by Qatar’s Ministry of Communications and Information Technology, the initiative seeks to advance sovereign AI development. The team released details of the platform and its underlying models in January 2025, publishing technical reports and making model weights available to the research community through platforms like Hugging Face.
Inside the model
The platform relies on two core language models. Fanar Star, with seven billion parameters, was trained from scratch on one trillion tokens. Fanar Prime, containing nine billion parameters, builds upon the Gemma-2 architecture using continuous pretraining on the same dataset. The training data includes roughly 410 billion Arabic tokens, 515 billion English tokens, and 102 billion code tokens. Post-training involved supervised fine-tuning and direct preference optimization to align outputs with Islamic values and Arab cultures.
What it can do
Fanar handles text, speech, and visual tasks. It includes bilingual speech recognition for multiple Arabic dialects such as Gulf, Levantine, and Egyptian. The system features specialized retrieval augmented generation tools for Islamic queries and recent events. Benchmark results indicate strong performance on Arabic-specific tests, often surpassing comparable models in cultural alignment and language understanding. It also provides an attribution service to verify the authenticity of generated facts.
How to use it
Model weights for Fanar-1-9B-Instruct are available openly via Hugging Face under permissive terms for research and development. The full platform, including advanced multimodal features and API access, is managed through the Fanar website. Developers can request API keys to integrate the models into their applications. The interface supports both direct model deployment and cloud-based interaction through Fanar Chat.
Safety and testing
The developers aligned the models with Islamic values and Arab cultural norms during training. They advise against using the system for high-stakes decisions in legal, medical, or financial domains. While extensive testing was conducted, the creators note that the model may still produce biased or incorrect outputs. Users are encouraged to implement additional safety checks and domain-specific fine-tuning for sensitive applications.
Why it matters
Fanar represents a significant step in developing sovereign AI infrastructure for the Middle East. By focusing on Arabic dialects and cultural specificity, it addresses gaps left by general-purpose global models. The release of open weights encourages further research into Arabic natural language processing. It demonstrates how regional institutions can build competitive generative AI systems tailored to local linguistic and ethical requirements.
Questions people ask about Fanar
What languages does Fanar support?
Fanar primarily supports Arabic and English. It handles Modern Standard Arabic and several dialects, including Gulf, Levantine, and Egyptian. The models are designed to be bilingual, allowing for seamless interaction in either language or a mix of both.
Is Fanar open source?
The model weights for Fanar-1-9B-Instruct are available on Hugging Face, allowing developers to download and run them locally. However, the full multimodal platform and certain advanced features are accessed via an API provided by the Qatar Computing Research Institute.
What makes Fanar different from other LLMs?
Fanar is specifically optimized for Arabic language and culture. It includes features like Islamic retrieval augmented generation and dialect-specific speech recognition. The training data and alignment processes were curated to reflect Arab cultural values and religious contexts.
Good reading and sources
There is no English Wikipedia article on Fanar yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2501.13944, huggingface.co/QCRI/Fanar-1-9B-Instruct. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.