AI models tested for Arabic

Arabic sentence-embedding model built on AraBERT v0.2 with Matryoshka representation learning, trained on Arabic NLI triplets so a single model serves several embedding dimensions. Widely used for Arabic semantic search and RAG retrieval, and one of the few Arabic-first embedding models with a Saudi research affiliation.

Frontier Arabic model announced by HUMAIN at LEAP Riyadh on 3 Sept 2026: a 428B-parameter mixture-of-experts model on the MiniMax-M3 lineage, further pretrained on more than one trillion tokens of Arabic-native content. Commissioned by HUMAIN and delivered by MiniMax. As of 10 Sept 2026 it is research preview only — reachable through HUMAIN Node's playground and an OpenAI-compatible endpoint, with no weights published. HUMAIN says it expects to release weights under the MiniMax Community License after safety training, targeted for October 2026.

Arabic sentence-embedding model producing 768-dimension vectors, trained on Arabic natural-language-inference and semantic-similarity data over AraBERTv02. Developed with support from Prince Sultan University in Riyadh, and the most used Arabic embedding model with Saudi provenance. Apache-2.0.

Small native-Arabic GPT models for research from the Saudi ArabianLLM initiative, published at 0.1B, 0.3B, 0.8B and 1.5B with task-tuned QA, summarisation and sentiment variants. Research-scale rather than production: the flagship repo has not changed since Feb 2024.

Arabic-localised LLM family with cultural alignment, co-developed by KAUST and King Abdulaziz University with CUHK-Shenzhen and SRIBD. Version 2 ships base and chat variants at 8B, 32B and 70B. The repos have not been updated since Nov 2024 and there is no v3.

Cross-dialect Arabic translation system from Misraj in Khobar, fine-tuned from Kuwain-1.5B in two variants: Lahjawi-D2D translates between 15 Arabic dialects and Lahjawi-D2MSA converts any dialect into MSA. The paper names Egyptian, Emirati, Jordanian, Palestinian, Levantine and Maghrebi among the dialects it evaluates but does not enumerate all 15. No weights are published.

Compact bidirectional Arabic–English translation model fine-tuned from Kuwain-1.5B; reported state of the art on the Tarjama-25 benchmark, which Misraj also published. Weights are not on Hugging Face — Misraj released the Tarjama-25 dataset and the evaluation code but keeps the model itself behind its Kawn platform.

Small Arabic/English model built in Khobar by injecting Arabic into an existing English model rather than pretraining from scratch; it is the base for both Mutarjim and Lahjawi. Misraj publishes its datasets and evaluation code openly but has released no Kuwain weights — the Hugging Face Misraj org hosts datasets only, and the model is reached through Misraj's Kawn platform.

Compact model tuned for Arabic/English RAG question-answering and entity extraction. Gemma-derived, so it ships under the Gemma Terms of Use; GGUF and 4-bit builds are published alongside it.

Gemma-based Arabic model that topped open Arabic rankings, outperforming much larger models. Because it is built on Gemma it inherits Google's Gemma Terms of Use rather than a standard open-source licence.

Largest model in the ALLaM family. SDAIA introduced the ALLaM family; HUMAIN adopted the 34B model and deployed it as the closed HUMAIN Chat service (Aug 2025). No weights, technical report or public API have been released, so there is no repository to link. An independent UI-level study evaluated it through HUMAIN Chat across MSA, five regional dialects and code-switching, but does not name those dialects.

Saudi national Arabic LLM built by the National Center for AI at SDAIA — 7B open weights, trained on 4T English tokens then 1.2T mixed Arabic/English tokens. The original ALLaM-AI Hugging Face org now redirects to humain-ai, so the weights resolve under HUMAIN's namespace; the model card still credits NCAI/SDAIA as the developer. The card claims MSA and English only.

Bilingual Arabic-English multimodal model built on Qwen2-VL and trained on 3.6M multimodal samples, roughly a third of them authentic Arabic. Strong on Arabic OCR and document understanding, with human evaluation across 17 Arab countries. MIT-licensed.

Whisper-base fine-tuned for Quranic Arabic recitation, reporting 5.75% WER on its evaluation set. Narrow by design — it targets recitation rather than conversational Arabic — but it is the most-adopted open Arabic speech model outside the general Whisper checkpoints, and underpins Tarteel's memorisation app. A tiny variant is also published.

2.35B Arabic-first speech recognition trained on 300k+ hours, naming Gulf, Egyptian, Levantine and Maghrebi coverage plus Arabic-English code-switching. Its card claims the top place on the Open Universal Arabic ASR Leaderboard — a leaderboard run by Elm, a Saudi company. Custom AudarAI Community Licence, not open source.

Open LLM family. The only model in this directory whose developer explicitly names a Saudi dialect: Qwen documents 119 languages and dialects including Arabic (Standard, Najdi, Levantine, Egyptian and others). Apache-2.0.

Arabic ASR trained with weak supervision on 30K+ hours; the paper claims best-in-class accuracy across 18 dialects. No weights are published — there is no CNTXT organisation or Munsit repository on Hugging Face — and the model is commercial API-only, so the dialect claims cannot be independently checked.

BERT models pre-trained per Arabic variant (MSA, dialectal, classical) for NER, POS, sentiment and dialect identification. The mix checkpoint remains the most-used entry point despite dating from 2021.

Arabic GPT-2 trained on the same corpus as AraBERTv2 and still a common Arabic generation baseline, published in base, medium, large and mega sizes. No licence is stated on the base model card.

The pioneer Arabic BERT encoder and still the standard baseline in Arabic NLP. The original bert-base-arabert card now flags itself as superseded, so this row points at AraBERT v0.2 base — the most-downloaded variant, trained on 77GB of Arabic text. No licence is stated on the model card or the aub-mind/arabert GitHub repo.

Arabic-centric flagship of the Fanar 2.0 release (Mar 2026), continually pretrained from google/gemma-3-27b-pt on ~166B Arabic, English and code tokens with 32K context. It adds native Arabic reasoning traces, selective thinking mode and tool calling. This repo is text-in/text-out; image generation, image understanding and poetry are separate Fanar-2 models.

Qatar's sovereign Arabic LLM. This 8.7B instruct model — the 'Prime' branch — continually pretrains google/gemma-2-9b on 1T Arabic and English tokens; a separate 7B 'Star' model was trained from scratch. The card claims MSA plus Gulf, Levantine and Egyptian dialects, and alignment with Islamic values and Arab culture.

First Arabic model in the Falcon series (May 2025), built on Falcon3-7B by extending the tokenizer with 32,000 Arabic-specific tokens and training on non-translated Arabic data. TII kept it closed: no weights are published under tiiuae — only the evaluation detail datasets — and the model is reachable through chat.falconllm.tii.ae. The announcement claims MSA plus unnamed 'key dialects'.

Successor to the Jais family, released as open weights in Aug 2026 in 8B and 70B chat sizes (plus GGUF builds). The card states it covers MSA, regional dialects and Arabic–English code-switching without naming the dialects. Repos are gated behind a contact-sharing agreement.

Landmark open Arabic LLM family (590M–70B), trained on Cerebras Condor Galaxy. The inceptionai org was renamed inception42, so the old repo URL only resolves via redirect; the repo is gated behind a contact-sharing agreement and has not been updated since Sept 2024. Superseded in practice by Jais 2.