Jais-2-8B-Chat: benchmark results
G42 Inception, Cerebras and MBZUAI's Jais 2 series chat model at 8B parameters, released under an Apache 2.0 license. Provider: Inception (G42), Cerebras & MBZUAI. Released 2025-12-09. Access: Open.
Aggregate Rating 1493 ± 19, rank #802 of 1635 rated models, from 58 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| QIMMA - ArabicMMLU | 73.96 | Accuracy (%, log-likelihood multiple choice) | 93.9 |
| QIMMA - PALMX | 78.98 | Accuracy (%, log-likelihood multiple choice) | 92.7 |
| QIMMA - MizanQA | 61.33 | Normalized multiple-choice probability score (%) | 86.6 |
| QIMMA - AraTrust | 87.55 | Accuracy (%, log-likelihood multiple choice) | 84.8 |
| QIMMA - ArabCulture | 68.99 | Accuracy (%, log-likelihood multiple choice) | 84.1 |
| QIMMA - Overall | 59.33 | Mean score across 14 benchmarks (%) | 78 |
| ArabCulture-Dialogue - Dialect Steering | 20.8 | Dialect identity accuracy (%): share of continuations whose | 72.7 |
| ArabCulture-Dialogue - Dialect Steering Quality | 73 | Continuation quality (0-100): GPT-5 judge rating on 1-5 resc | 72.7 |
| ArabCulture-Dialogue - MCQ with Country Context (Dialect) | 73.1 | Accuracy (%) selecting the one culturally appropriate contin | 72.7 |
| QIMMA - MedAraBench | 41.25 | Accuracy (%, log-likelihood multiple choice) | 72 |
| QIMMA - AraDiCE-Culture | 65.56 | Accuracy (%, log-likelihood multiple choice) | 69.5 |
| QIMMA - FannOrFlop | 51.94 | F1 (%) | 69.5 |
Interactive version: theaggregate.ai/model?slug=jais-2-8b-chat · How It Works · Data refreshed daily, snapshot 2026-11-05.