Arabic language
Understanding and answering in Arabic
Najd Arena
Loading evaluation results…
Independent evaluation by Najd Research
The benchmark for
Arabic and Saudi AI.
Understand model performance where language,
local knowledge, and real-world tasks meet.
Model comparison
One execution path and thinking setting at a time.
| Model | Acceptable answer rate | Acceptable / total | Technical failures | |
|---|---|---|---|---|
 HUMAIN M3HUMAIN | 67.35% | 4,101 / 6,089 | 0 | |
MiniMax M3MiniMax | 66.12% | 4,026 / 6,089 | 10 |
Correct + possible correct, divided by all outputs. Technical failures contribute zero. Differences are descriptive, not a significance claim. Historical evaluation and publication dates have not been verified.
Performance highlights
Acceptable answer rate · Historical results · Higher is better
Saudi knowledge
Knowledge grounded in Saudi context
Tool decisions
Choosing tools for the task
What is behind the score?
Answer labels across all tasks · Direct (Raw) · default thinking
CorrectPossible correctPartial correctWrongTechnical failure
Select a segment to inspect its count and share.
Label definitions and provenance
Correct: normalized gold answer found. Possible correct: valid alternative. Partial correct: material omission or error. Wrong: neither a valid answer nor meaningful correct part. Technical failures remain a separate segment. Labels come from the preserved September 9 snapshot, verified against the Arena seed; they have not received a full-answer audit.
Find the right modelStart with your use case
Build your benchmarkEvaluate what matters to your team
Latest updatesFollow new results and platform changes
Explore thinking levels
Compare every recorded setting. Direct and Pi execution remain separate.
Compare model performance
Historical results Acceptable output rate · Higher is better
defaultHUMAIN M367.35%
MiniMax M366.12%
HUMAIN M367.35%Historical results · Evaluation and publication dates have not been verified. Evidence policy →
Fixed 0–100% scale · Failures count as zero.
How to read this chart
Raw is direct model execution; Pi uses the agent runtime. These paths and their prompt settings are kept separate. Levels are API settings, not verified measures of thinking effort. “Default” uses provider defaults; “disabled” requests no reasoning. Settings may behave differently across providers.
Each overall configuration covers 6,089 case IDs. These historical pre-audit results include all cases. Earlier case contents differ from the current release; no significance test is available. Hover, focus, or tap a bar to inspect its counts.
Built for the questions that matter hereBring your model.
Request an evaluation