Skip to content
Living editionSources checked 16 November 2026. We re-check and add results as new Arabic benchmarks are published.

Research · Model scoreboard

Model scoreboard

How each AI model performs in Arabic: every published score on an Arabic benchmark, where the model ranks against the others measured on the same test, and how it compares with models of its own size. The scores come from our Arabic benchmark map.

Compare models

How each model performs on Arabic benchmarks

Pick a company and a model to see every Arabic test it has a published score on, how it scored, and where it ranks against the other models measured on the same test. Rankings only compare models tested in the same paper or leaderboard, so each rank also shows who else was in the field and where the model places among models of its own size. Select a benchmark to see its details and every model’s score.

Gemini 3 Flash

Google · Closed lab model
17tests scored, 4 benchmarks
12tests where it ranks first
15 of 15tests where it is in the top half
1st of 15top score1st of 10 closed lab modelsField: 10 closed, 2 mid, 3 small
1st of 14top score1st of 7 closed lab modelsField: 7 closed, 3 mid, 4 small
Only modelno peers measured
1st of 15top score1st of 10 closed lab modelsField: 10 closed, 2 mid, 3 small
1st of 15top score1st of 10 closed lab modelsField: 10 closed, 2 mid, 3 small
1st of 14top score1st of 7 closed lab modelsField: 7 closed, 3 mid, 4 small
1st of 14top score1st of 7 closed lab modelsField: 7 closed, 3 mid, 4 small
3rd of 14best: Claude Opus 4.5 91.8%3rd of 7 closed lab modelsField: 7 closed, 3 mid, 4 small
1st of 15top score1st of 10 closed lab modelsField: 10 closed, 2 mid, 3 small
1st of 4top score1st of 2 closed lab modelsField: 2 closed, 1 large, 1 mid
1st of 5top scoreField: 1 closed, 2 large, 2 mid
Only modelno peers measured
1st of 5top scoreField: 1 closed, 2 large, 2 mid
1st of 14top score1st of 7 closed lab modelsField: 7 closed, 3 mid, 4 small
3rd of 14best: Claude Sonnet 4.5 78.4%3rd of 7 closed lab modelsField: 7 closed, 3 mid, 4 small
6th of 14best: GPT-4o 78.1%5th of 7 closed lab modelsField: 7 closed, 3 mid, 4 small
1st of 2top scoreField: 1 closed, 1 mid
Gemini 3 FlashOther models on the same testScores shown as % of each test’s maximum; error-rate tests as 100 minus the error rate
123 models
OpenAI19 models
Google21 models
Anthropic13 models
Alibaba (Qwen)14 models
QCRI (Fanar)6 models
SDAIA / HUMAIN (ALLaM)4 models
Inception, G42 (Jais)6 models
Meta8 models
SILMA AI1 model
HUMAIN2 models
Cohere4 models
TII (Falcon)6 models
Arabic.AI1 model
IFM (MBZUAI) / Inception (Jais)1 model
Mistral1 model
Ansari (Islamic assistant)1 model
DeepSeek4 models
FreedomIntelligence (AceGPT)3 models
NAMAA Community1 model
Navid AI1 model
University of Helsinki1 model
Academic (MetaMath)1 model
Inception (G42)1 model
Tuwaiq Academy1 model
MBZUAI (Nile-Chat)1 model
NVIDIA1 model
Reading the scores

What the scores are and are not

Scores come from different papers, harnesses, prompts and dates, so they are not a controlled comparison. Bar colors separate closed models from global labs, Arabic-first and regional models, and open-weight global models. Error-rate tests (speech recognition) are shown as 100 minus the error rate. Ratings on a 1–5 or 0–10 scale are shown as a share of the maximum. The benchmark map’s appendix defines every color and tag.

Co-author a benchmark with us

Most Arabic domains are still unmeasured. Have a domain where Arabic performance matters? We’ll build the benchmark with your team and our native experts.

Talk to our team