Najd Arena · By Najd Research

AI performance.
Measured in context.

Choosing a model starts with the problems it needs to solve. Najd Arena makes Arabic and Saudi AI performance easier to compare, understand, and reproduce.

Explore model results →Talk to Najd ↗

Our purpose

Set a clearer standard
for relevant AI evaluation.

General benchmarks offer part of the picture. Arabic language, local knowledge, and business workflows deserve their own evidence.

We are building a reference for people choosing models for Arabic customer support, Saudi knowledge questions, document assistants, and tool-using agents.

Evaluation coverage

What we measure

Read the methodology →

Language & local context

Arabic understanding, Saudi knowledge, and culturally relevant questions.

Available results ↗02

Documents & retrieval

Answering questions from supplied information, with task-level breakdowns.

Available text results ↗03

Coding with a harness

Models inside Pi on the available coding tasks, keeping execution separate from direct model calls.

Limited task coverage ↗04

Enterprise vision

Visual understanding, Arabic OCR, document extraction, charts, and tables.

In development ↗

Evidence you can inspect

Comparable settings

Direct model calls and agent-harness runs stay separate. Thinking settings remain visible.

Transparent scoring

Inspect task scores, counts, and failure handling. A small score difference is not a claim of statistical significance.

Reproducible foundations

Dataset versions and evaluation methods provide the context needed to interpret each result.

The currently imported results come from a preserved historical evaluation. They are not Najd-reviewed runs on the exact current dataset. Each result retains its scope and limitations.