Arabic LLM Models
The ecosystem of Arabic LLM models is quickly expanding, presenting a challenge in keeping up with the latest developments. This article aims to address this issue by serving as a comprehensive resource that continuously updates with new Arabic LLM models, providing users with the necessary information and links to choose the best model for their specific task. This living document will serve as a go-to source for all your Arabic LLM model needs.
Selection Criteria
For a model to be included, one of the following needs to be true
- Model is open source
- Model can be tried online via a link
- Model is offered as an API
General Purpose Models
Below is a list of general purpose Arabic Models (Order does not indicate performance)
| Name | Size | License | Link | Comments |
|---|---|---|---|---|
| SILMA v1.0 | 9B | Open-weight (Gemma) | https://huggingface.co/silma-ai/SILMA-9B-Instruct-v1.0 | Based on Gemma. Was #1 on OALL V1 benchmark |
| Fanar | 7B | Closed | https://chat.fanar.qa/ | Qatar's Sovereign Model |
| Allam | 7B | Open Weight (Apache 2.0) | https://huggingface.co/ALLaM-AI/ALLaM-7B-Instruct-preview | Saudi's Sovereign Model |
| Jais | 590M to 70B | Open Weight (Apache 2.0) | https://huggingface.co/collections/inceptionai/jais-family-66add8bb9c381f5492ddb6f4 | UAE's Arabic Models and one of the first players |
| AceGPT-7B-chat | 7B-32B | Open Weight (Apache 2.0) | https://huggingface.co/FreedomIntelligence/AceGPT-7B-chat | |
| Cohere command-r7b-arabic | 8B | Open Weight (CC Non Commercial 4.0) | https://huggingface.co/CohereForAI/c4ai-command-r7b-arabic-02-2025 | General purpose + optimized for RAG |
| Cohere aya-expanse | 8B-32B | Open Weight (CC Non Commercial 4.0) | https://huggingface.co/CohereForAI/aya-expanse-32b | |
| Gemma | 2B-27B | Open Weight (Gemma) | https://huggingface.co/google/gemma-2-9b-it | Google's multilingual open model which includes Arabic |
| Qwen 2.5 | 0.5B-72B | Open Weight (Apache 2.0) | https://huggingface.co/Qwen/Qwen2.5-0.5B | Alibaba's multilingual open model which includes Arabic |
| Llama 3.3 | 70B | Open Weight (Llama 3.3 Community License Agreement) | https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct | Meta's multilingual open model which includes Arabic. Very good performance in OALL benchmark |
| Llama 3.2 | 1B-3B | Open Weight (Llama 3.3 Community License Agreement) | https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct | Meta's multilingual open model which includes Arabic. |
| Phi 3.5 | 4B | Open Weight (MIT) | https://huggingface.co/microsoft/Phi-3.5-mini-instruct | Microsoft's multilingual open model which includes Arabic. |
| Phi 4 | 4B | Open Weight (MIT) | https://huggingface.co/microsoft/Phi-4-mini-instruct | Microsoft's multilingual open model which includes Arabic. |
| Mistral Saba | 24B | Closed | https://mistral.ai/news/mistral-saba | Offered via API only |
| Ar-stablelm-2-chat | 1.6B | Open Weight (MIT) | https://huggingface.co/stabilityai/ar-stablelm-2-chat | |
| Yehia-7B-preview | 7B | Open Weight (MIT) | https://huggingface.co/Navid-AI/Yehia-7B-preview | Based on Allam |
RAG-optimized Models
Below is a list of models trained and optimized for RAG Generatation use-cases
| Name | Size | License | Link | Comments |
|---|---|---|---|---|
| SILMA Kashif v1.0 | 2B | Open-weight (Gemma) | https://huggingface.co/silma-ai/SILMA-Kashif-2B-Instruct-v1.0 | Benchmark |
| Cohere command-r7b-arabic | 8B | Open Weight (CC Non Commercial 4.0) | https://huggingface.co/CohereForAI/c4ai-command-r7b-arabic-02-2025 | General purpose + optimized for RAG |
Vision & OCR
Below is a list of models with multimodal capabilities (Vision, text, ...etc)
| Name | Size | License | Link | Comments |
|---|---|---|---|---|
| AIN | 8B | Open Weight (MIT) | https://huggingface.co/MBZUAI/AIN | Based on Qwen |
| Qari OCR | 2B | Open Weight (Apache 2.0) | https://huggingface.co/NAMAA-Space/Qari-OCR-0.1-VL-2B-Instruct | Based on Qwen. OCR only |
| Cohere aya-vision | 8B-32B | Open Weight (CC Non Commercial 4.0) | https://huggingface.co/collections/CohereForAI/c4ai-aya-vision-67c4ccd395ca064308ee1484 |
Dialect-optimized Models - Syrian Arabic
Models optimized for the Leventian dialect
| Name | Size | License | Link | Comments |
|---|---|---|---|---|
| Shahin-v0.1 | 14B | Open Weight (Apache 2.0) | https://huggingface.co/malhajar/Shahin-v0.1 | Based on Qwen |
Dialect-optimized Models - Morocan Arabic
Models tuned for Darija, the colloquial Arabic of Morocco
| Name | Size | License | Link | Comments |
|---|---|---|---|---|
| Atlas-Chat | 9B-27B | Open Weight (Gemma) | https://huggingface.co/MBZUAI-Paris/Atlas-Chat-9B | based on Gemma |
Dialect-optimized Models - Tunisian Arabic
Models tuned for Tunisian Arabic
| Name | Size | License | Link | Comments |
|---|---|---|---|---|
| Labess Chat | 7B | Open Weight (apache-2.0) | https://huggingface.co/linagora/Labess-7b-chat | based on Jais |
A model is missing?
If you believe that a model is not included in the list, please leave a comment below. If it meets the necessary criteria, it will be added.
How to choose a model?
In addition to testing the model on real-world use cases, benchmarks are valuable for evaluating various aspects of the model's performance.
The following post contains a list of Arabic AI Benchmarks https://huggingface.co/blog/silma-ai/arabic-ai-benchmarks-and-leaderboards
