A field guide to the models shaping AI across language, code, image, video, audio, and the enterprise. It covers what each one genuinely leads at, where it still falls short, where it comes from, and when it shipped.
Two independent studies see different leaders because they measure different behavior. Microsoft puts the UAE first for estimated generative-AI use; Cybernews puts it third for mobile AI-app downloads. The percentages are kept separate rather than averaged.
AI user share · June 2026
AI-app downloads per population · 2025
These rankings are not combined: their populations, periods and observed behavior differ.
40 models have enough independent measurement to be scored. Each category is marked on its own rubric, so this mixed view is a rough guide — use the tabs to compare like with like.
Only the top is published. Further down a ranking the marks are mostly older models kept for reference, or models with just enough measurement to be scored, so a “lowest scored” table would say more about the gaps in the evidence than about the products. Each model’s own page shows its marks and what’s still missing.
The score measures capability only: price and availability are measured too, but shown on each model’s page rather than folded into this number. Marks use the low end of each source’s confidence interval, and models whose intervals overlap are marked as tied, because the evidence can’t separate them. Models with too little independent measurement are under evaluation and don’t appear here at all. How scoring works
Ranked on how well they handle Arabic — which the scores above say nothing about, because the benchmarks behind them are run in English. Two independent leaderboards publish an Arabic measurement in a form that can be read and checked. Both are shown for every model, and both are combined into the figure the table is ordered by.
AraGen v3 · 67 · Apr 4, 2026OALL v2 · 163 · Mar 2, 2026
The large numbers are each board’s own result, on its own scale. The table is ordered by the combined figure beside them, which is why a model can show a high number here and still sit low: the combined figure carries the cost of converting between the two scales, and the row says which board measured it.
MBZUAI
CerebrasInception · UAEInception's flagship, trained with MBZUAI and Cerebras on Arabic and English, with a vocabulary built for Modern Standard Arabic, regional dialects and Arabic–English code-switching.The two scales aren’t interchangeable, so they’re not simply averaged. A line is fitted across the 25 models both boards have run, where the agreement is moderate rather than tight (r = 0.67). Converting an OALL result into AraGen’s terms costs ±2.2 marks at 95%, and the two figures are then weighted by how firm each one is, so a directly measured score counts far more than a converted one.
A model only one board has measured therefore carries a wide band, is marked at the low end of it like every other mark on the site, and ties with much of the table. That is the honest reading rather than a fault in the table: its Arabic is as good as the board that measured it says, and what that means on the other scale isn’t known. Read the two source columns, not just the combined figure — where they disagree, the disagreement is the finding.
Each board’s own figure follows the same rule as the rest of the site: a published percentage divided by 10, never an estimate. These models are kept apart from the directory’s 100 because they are judged on different benchmarks, against a different field — a place here says nothing about where a model stands globally, and a place there says nothing about its Arabic. How scoring works
LMArena splits the same blind, side-by-side votes by the language of the prompt. This is the directory’s models on the prompts that weren’t in English — the same voters, the same models, different work. No arena breaks Arabic out on its own, which is why the board above is built from Arabic benchmarks instead.
LMArena406 models measured, published Oct 2, 2026
28 of the directory’s models clear the vote floor on this board. The top 10 are listed, highest mark first.
The two places are counted in the same field of models, so they can be read against each other — but read them gently. Around any given place the ratings sit inside one another’s confidence bands, so a move of even twenty places usually can’t be told apart from noise. A model far from its English place is worth a look; a few places either way is not.
Marked on the arena’s own fixed scale at the low end of each published interval, the same rule the directory’s scores follow, and a rating built on fewer than 500 votes is treated as unmeasured rather than weak. How scoring works
Each maker’s headquarters country. The frontier is concentrated in two places, but Europe, the Gulf, Israel, and Asia-Pacific each ship models worth knowing.