Model Overview
Model Features
Model Capabilities
Use Cases
🚀 Fanar-1-9B-Instruct
Fanar-1-9B-Instruct is a potent Arabic-English large language model (LLM) developed by the Qatar Computing Research Institute (QCRI) at Hamad Bin Khalifa University (HBKU), a member of the Qatar Foundation for Education, Science, and Community Development. It is the instruction-tuned version of Fanar-1-9B. We continuously pretrain the google/gemma-2-9b
model on 1 trillion Arabic and English tokens. Special attention is paid to the richness of the Arabic language by supporting Modern Standard Arabic (MSA) and a diverse range of Arabic dialects, such as Gulf, Levantine, and Egyptian. Through meticulous curation of pretraining and instruction-tuning data, Fanar models are aligned with Islamic values and Arab cultures.
Fanar-1-9B-Instruct is a core part of the Fanar GenAI platform, which offers a variety of capabilities, including image generation, video and image understanding, deep thinking, advanced text-to-speech (TTS) and automatic-speech-recognition (ASR), attribution and fact-checking, Islamic RAG, and several other features.
We have published a comprehensive report with all the details about our Fanar GenAI platform. We also provide an API to our models and the GenAI platform (request access here).
🚀 Quick Start
Fanar-1-9B-Instruct is compatible with the Hugging Face transformers
library (≥ v4.40.0). Here's how to load and use the model:
from transformers import AutoTokenizer, AutoModelForCausalLM
model_name = "QCRI/Fanar-1-9B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
# message content may be in Arabic or English
messages = [
{"role": "user", "content": "ما هي عاصمة قطر؟"},
]
inputs = tokenizer.apply_chat_template(messages, tokenize=False, return_tensors="pt")
outputs = model.generate(**tokenizer(inputs, return_tensors="pt", return_token_type_ids=False), max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Inference using VLLM is also supported:
from vllm import LLM, SamplingParams
model_name = "QCRI/Fanar-1-9B-Instruct"
llm = LLM(model=model_name)
sampling_params = SamplingParams(temperature=0.7, max_tokens=256)
# message content may be in Arabic or English
messages = [
{"role": "user", "content": "ما هي عاصمة قطر؟"},
]
outputs = llm.chat(messages, sampling_params)
print(outputs[0].outputs[0].text)
✨ Features
- Multilingual Support: Supports both Arabic and English, with a focus on the richness of the Arabic language, including Modern Standard Arabic (MSA) and various dialects.
- Cultural Alignment: Aligned with Islamic values and Arab cultures through meticulous data curation.
- Comprehensive Capabilities: Part of the Fanar GenAI platform with a suite of features such as image generation, video and image understanding, etc.
📦 Installation
The model can be installed using the Hugging Face transformers
library. Make sure you have transformers
version ≥ v4.40.0 installed. You can install it using the following command:
pip install transformers>=4.40.0
📚 Documentation
We have published a comprehensive report with all the details regarding our Fanar GenAI platform. We also provide an API to our models and the GenAI platform (request access here).
🔧 Technical Details
Model Details
Property | Details |
---|---|
Developed by | QCRI at HBKU |
Sponsored by | Ministry of Communications and Information Technology, State of Qatar |
Model Type | Autoregressive Transformer |
Parameter Count | 8.7 Billion |
Context Length | 4096 Tokens |
Input | Text only |
Output | Text only |
Training Framework | LitGPT |
Pretraining Token Count | 1 Trillion (ar + en) |
SFT Instructions | 4.5M |
DPO Preference Pairs | 250K |
Languages | Arabic, English |
License | Apache 2.0 |
Model Training
Pretraining
Fanar-1-9B-Instruct was continually pretrained on 1T tokens, with a balanced focus on Arabic and English: ~515B English tokens from a carefully curated subset of the Dolma dataset, 410B Arabic tokens that we collected, parsed, and filtered from a variety of sources, and 102B code tokens curated from The Stack dataset. Our codebase used the LitGPT framework.
Post-training
Fanar-1-9B-Instruct underwent a two-phase post-training pipeline:
Phase | Size |
---|---|
Supervised Fine-tuning (SFT) | 4.5M Instructions |
Direct Preference Optimization (DPO) | 250K Preference Pairs |
📄 License
This model is licensed under the Apache 2.0 License.
💻 Usage Examples
Basic Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
model_name = "QCRI/Fanar-1-9B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
# message content may be in Arabic or English
messages = [
{"role": "user", "content": "ما هي عاصمة قطر؟"},
]
inputs = tokenizer.apply_chat_template(messages, tokenize=False, return_tensors="pt")
outputs = model.generate(**tokenizer(inputs, return_tensors="pt", return_token_type_ids=False), max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Advanced Usage
from vllm import LLM, SamplingParams
model_name = "QCRI/Fanar-1-9B-Instruct"
llm = LLM(model=model_name)
sampling_params = SamplingParams(temperature=0.7, max_tokens=256)
# message content may be in Arabic or English
messages = [
{"role": "user", "content": "ما هي عاصمة قطر؟"},
]
outputs = llm.chat(messages, sampling_params)
print(outputs[0].outputs[0].text)
Intended Use
Fanar-1-9B-Instruct is built for:
- Conversational agents (Arabic only or bilingual)
- Cultural and dialectal question answering in Arabic
- Educational, governmental, and civic NLP applications focused on the Arab world or Arabic-speaking audiences
- Research on Arabic natural language generation and understanding
Fanar-1-9B-Instruct can be deployed as part of a broader AI system. Developers are encouraged to implement proper safeguards to ensure culturally respectful, accurate, and safe deployment. It should not be used to generate or spread harmful, illegal, or misleading content.
A version of this model can be accessed through Fanar Chat. We are continuously improving the Fanar’s models and capabilities, and answers can differ from what you get from Fanar-1-9B-Instruct.
Ethical Considerations & Limitations
Fanar-1-9B-Instruct is capable of generating fluent and contextually appropriate responses. However, as with any generative model there are uncertainties. The model may produce biased, offensive, or incorrect outputs. The model is not suitable for high-stakes decision-making (e.g., legal, medical, or financial advice). Though we have extensively tested Fanar-1-9B-Instruct and attempted to mitigate these issues, we cannot redress every possible scenario. Thus, we advise developers to implement safety checks and perform domain-specific fine-tuning for sensitive use cases. Kindly refer to our Terms of Service and Privacy Policy.
The output generated by this model is not considered a statement of QCRI, HBKU, Qatar Foundation, MCIT or any other organization or individual.
Evaluation
Evaluation was conducted using a modified version of the LM Evaluation Harness and internal cultural alignment benchmarks.
Model | MMLU (5-shot) | MMMLU (Arabic) (0-shot) | ArabicMMLU (3-shot) | HellaSwag (0-shot) | PIQA (0-shot) | ARC Challenge (0-shot) | Belebele (Arabic) (3-shot) | ACVA (5-shot) | GSM8k | OALL (0-shot) | OALL v2 (0-shot) | Almieyar Arabic (3-shot) | Arab Cultural MCQ (3-shot) | AraDiCE PIQA (MSA) (0-shot) | AraDiCE PIQA(Egy) (0-shot) | AraDiCE PIQA(Lev) (0-shot) | AraDiCE ArabicMMLU(Egy) (0-shot) | AraDiCE ArabicMMLU(Lev) (0-shot) |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Fanar-1-9B-it | 71.53% | 58.89% | 67.69% | 83.16% | 82.54% | 67.15% | 83.22% | 80.02% | 74.60% | 68.32% | 66.29% | 78.68% | 72.40% | 67.68% | 63.66% | 59.03% | 59.63% | 60.62% |
ALLaM-7B-Instruct-preview | 60.72% | 54.89% | 68.59% | 76.35% | 80.52% | 51.62% | 75.80% | 74.52% | 46.63% | 57.31% | 63.66% | 76.31% | 74.20% | 67.52% | 63.44% | 60.88% | 62.50% | 64.17% |
aya-expanse-8b | 62.85% | 47.14% | 60.10% | 78.54% | 81.18% | 56.40% | 70.78% | 77.11% | 8.26% | 53.18% | 59.74% | 70.20% | 67.30% | 63.00% | 59.41% | 56.53% | 53.52% | 53.71% |
c4ai-command-r7b-arabic-02-2025 | 66.91% | 49.54% | 63.06% | 74.67% | 78.02% | 49.15% | 72.78% | 79.80% | 30.33% | 49.38% | 64.44% | 73.82% | 69.20% | 62.30% | 60.99% | 56.69% | 54.78% | 56.06% |
AceGPT-v2-8B-Chat | 66.45% | 51.16% | 62.61% | 79.21% | 80.58% | 53.50% | 74.56% | 77.66% | 41.77% | 50.16% | 60.40% | 74.31% | 68.90% | 64.58% | 61.32% | 56.91% | 54.53% | 53.91% |
gemma-2-9b-it | 71.65% | 57.93% | 64.16% | 79.06% | 79.38% | 63.99% | 78.31% | 80.67% | 60.95% | 56.11% | 64.21% | 73.69% | 68.60% | 61.26% | 59.96% | 57.24% | 57.95% | 59.25% |
jais-adapted-13b-chat | 56.64% | 44.45% | 58.97% | 80.86% | 80.47% | 54.27% | 67.52% | 75.24% | 44.05% | 46.41% | 56.56% | 65.46% | 65.30% | 61.10% | 58.05% | 55.77% | 52.87% | 53.59% |
jais-family-6p7b-chat | 49.42% | 41.59% | 55.80% | 72.04% | 74.05% | 44.62% | 65.11% | 72.04% | 53.68% | 48.20% | 54.73% | 61.72% | 64.10% | 62.51% | 60.12% | 57.24% | 49.11% | 47.49% |
Llama-3.1-8B-Instruct | 68.04% | 47.58% | 59.05% | 79.22% | 80.74% | 55.29% | 66.72% | 76.67% | 29.26% | 47.81% | 55.97% | 69.70% | 66.10% | 58.11% | 55.39% | 54.24% | 46.86% | 47.52% |
Qwen2.5-7B-Instruct | 74.21% | 55.63% | 63.96% | 80.44% | 79.92% | 55.03% | 74.61% | 78.09% | 71.34% | 54.19% | 62.69% | 75.69% | 68.10% | 60.55% | 58.65% | 56.04% | 48.74% | 53.42% |
Citation
If you use Fanar-1-9B-Instruct or the Fanar GenAI system in your research or applications, please cite:
@misc{fanarllm2025,
title={Fanar: An Arabic-Centric Multimodal Generative AI Platform},
author={Fanar Team and Ummar Abbas and Mohammad Shahmeer Ahmad and Firoj Alam and Enes Altinisik and Ehsannedin Asgari and Yazan Boshmaf and Sabri Boughorbel and Sanjay Chawla and Shammur Chowdhury and Fahim Dalvi and Kareem Darwish and Nadir Durrani and Mohamed Elfeky and Ahmed Elmagarmid and Mohamed Eltabakh and Masoomali Fatehkia and Anastasios Fragkopoulos and Maram Hasanain and Majd Hawasly and Mus'ab Husaini and Soon-Gyo Jung and Ji Kim Lucas and Walid Magdy and Safa Messaoud and Abubakr Mohamed and Tasnim Mohiuddin and Basel Mousi and Hamdy Mubarak and Ahmad Musleh and Zan Naeem and Mourad Ouzzani and Dorde Popovic and Amin Sadeghi and Husrev Taha Sencar and Mohammed Shinoy and Omar Sinan and Yifan Zhang and Ahmed Ali and Yassine El Kheir and Xiaosong Ma and Chaoyi Ruan}},
year={2025},
url={https://arxiv.org/abs/2501.13944},
}
Acknowledgements
This project is from Qatar Computing Research Institute (QCRI) at Hamad Bin Khalifa University (HBKU), a member of Qatar Foundation. We thank our engineers, researchers, and support team for their efforts in advancing Arabic-centric large language models. Special thanks to the Ministry of Communications and Information Technology, State of Qatar for their continued support by providing the compute infrastructure through the Google Cloud Platform.

