The Arabic LLM leaderboard in 2026: Falcon, Jais, ALLaM & Fanar

The Arabic AI leaderboard quietly threw out a chunk of its own test.

The reason is the most important sentence in Arabic AI right now: the questions had been translated from English.

Translated tests smuggle in English phrasing, English assumptions, English culture. They end up measuring how well a model handles a translation — not how well it actually knows Arabic. So the community rebuilt the benchmark around questions written natively in Arabic, about Arabic and the region.²

That single move is the whole story of the Arabic LLM race in miniature: you cannot measure — or build — an Arabic model on data that was never really Arabic. Here's the current state of that race, and what's really driving it.

The scoreboard: the Open Arabic LLM Leaderboard

For years, "which Arabic model is best?" was answered with press releases. Now there's a public answer. The Open Arabic LLM Leaderboard (OALL), hosted on Hugging Face and started by the UAE's Technology Innovation Institute (TII), runs every submitted model through the same battery of Arabic tasks and ranks them.¹

Under the hood it uses Hugging Face's LightEval framework and scores models with normalized log-likelihood accuracy — a fair way to grade the multiple-choice and yes/no questions that make up most of the tests.²

What it tests changed in a way worth dwelling on. The first version (v1) mixed native Arabic datasets like AlGhafa and ACVA with Arabic translations of English benchmarks such as MMLU and EXAMS. Version 2 dropped the translations and moved to native benchmarks — ArabicMMLU, ALRAGE, AraTrust, MadinahQA — after community consensus that translated content wasn't a fair test of Arabic.²

The contenders: four Gulf models

The serious Arabic-first models come out of a Gulf sovereign-AI race — each backed by a state that decided owning its language model matters. They take different routes: some trained from scratch on Arabic, others adapt a strong multilingual base. Here's the field.

Falcon-H1 Arabic (TII, Abu Dhabi) is the current front-runner, released in 3B, 7B and 34B sizes.³ ALLaM is Saudi Arabia's national model — built by SDAIA with more than 400 experts and 160 government bodies, trained on 3T-plus tokens, now run by HUMAIN and powering the HUMAIN Chat app on its 34B model.⁴ Fanar is Qatar's sovereign model from QCRI at Hamad Bin Khalifa University; its 2.0 release is a 27B multimodal model handling text, image and speech.⁵ And Jais — from G42's Inception and MBZUAI — was one of the first major Arabic models trained from scratch, on 116B Arabic and 279B English tokens.⁶

Who leads right now

As of early 2026, the top of the board belongs to Falcon-H1 Arabic. Its 7B model averages 71.47% across the leaderboard's tasks — ahead of larger ~10B models including Fanar-1-9B and ALLaM 7B. Its 34B model scores 75.36%, ahead of some 70B-plus general models like Qwen2.5-72B and Llama-3.3-70B on Arabic.³

Two cautions before you read too much into any single number. First, it's a snapshot. New models and new versions land every few weeks; ALLaM was named a top Arabic model on ArabicMMLU, Falcon leads the aggregate today, and the order will shift again. Always check the live leaderboard for the current standings.

Second, an aggregate hides the split that matters to you. A model can top the average and still stumble on Gulf dialect, or on your industry's terms. Which is the next problem.

Why Arabic is uniquely hard to benchmark

English has one written standard. Arabic doesn't behave that way, and that's what makes both building and testing it hard.

• It's diglossic. Modern Standard Arabic (MSA) is the formal written register; almost no one speaks it day to day. People speak dialects — Gulf, Egyptian, Levantine, Maghrebi — that differ enough in words and grammar to trip a model trained mostly on MSA.
• It's morphologically rich. A single root spins out dozens of forms; short vowels are usually unwritten, so the same letters can be several different words until context resolves them.
• It's right-to-left, and mixes with Latin script and numerals in real text — a formatting minefield for tokenizers and annotators alike.
• Much of the knowledge is cultural. Names, history, religion, law and idiom that a translated benchmark simply misses.

This is exactly why native benchmarks matter. AraDiCE was built to test dialect and cultural competence across Egyptian, Levantine and Gulf Arabic — the things a translated MMLU can't see.⁷ And it's why researchers surveying the field keep landing on the same gap: the benchmarks, and the data behind them, are the bottleneck — not the model architecture.⁸

What actually decides the ranking

Strip the leaderboard down and here's what's left: every score is a measurement of data.

A model's Arabic ability comes from the Arabic examples it learned from. Its instruction-following comes from people writing and ranking Arabic prompts and answers — RLHF. And its score comes from human-written Arabic benchmarks. People at all three points. If those people don't speak the dialect, or miss the cultural context, the model learns the gap and the benchmark inherits it.

That's the throughline behind the leaderboard's own history. When the community dumped translated tests for native ones, it wasn't a modeling decision — it was a data-quality decision. Better data in, truer measurement out.

Which is where the work we do sits. The models above compete on capability; underneath all of them is a layer of Arabic annotation — labeling, transcription, dialect tagging, response ranking, evaluation — that decides how good any of them can get. The practical steps for improving an Arabic model rarely start with the architecture:

1. Gather representative Arabic data. Not just MSA — the dialects, domains and formats your users actually produce.
2. Label it with native speakers. Dialect, entities, intent, safety — annotated by people who hear the language the way your users do.
3. Rank answers for RLHF. Human preference on Arabic responses is what turns a base model into one that's actually helpful in Arabic.
4. Evaluate on your own data. A leaderboard tells you the general order; only your labeled examples tell you what wins on your task.

Who's teaching your machine?

I'm Ahmed Rafiq Fahmy. My co-founder and I ran data-annotation operations for a decade — both ex-Affectiva and ex-Smart Eye — as the customer, inside companies training AI, judging the tools and workforces from the buyer's seat. We built Annota8 into what that decade said the tool should have been, and made it Arabic-first from day one.

We don't build Arabic LLMs. We build the layer they're taught and tested on: the Annota8 Annotation Engine — 200 annotation UIs across 9 modalities, with Arabic and dialect workflows — plus an AI agent on your annotation data and one operating console to run the loop. Built in Saudi Arabia; our first paying customer is the University of California, San Diego.

Whichever model on the board you build on, its Arabic is only as good as the people who taught it.

Frequently asked