BALSAM - Factuality: leaderboard
BALSAM Arabic benchmark (SDAIA/KSAA, dataset version 2), Factuality category: judging the factual correctness of claims and answers against a provided text. The category has only multiple-choice tasks, so the site's overall score is 30% of the LLM-judged multiple-choice score and cannot exceed 30; higher is better.
Metric: Overall score (0-100, LLM-judged generation and multiple choice). Source: benchmarks.ksaa.gov.sa. Saturation forecast: Around February 2027. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Command A 03 2025 | 26.56 | #530 |
| 2 | Qwen 3 235B A22B | 25.67 | #323 |
| 3 | GLM-5 (Thinking) | 25.44 | #149 (GLM-5) |
| 4 | GLM-4.7 (Reasoning) | 25.44 | #202 (GLM-4.7) |
| 5 | Jais-2-70B-Chat | 25.44 | #430 |
| 6 | DeepSeek V3.2 | 24.78 | #210 |
| 7 | GPT-5.2 | 24.55 | #115 |
| 8 | Nova Pro (v1) | 24.33 | #508 |
| 9 | Claude Opus 4.5 | 23.11 | #87 |
| 10 | Fanar-C-2-27B | 22.11 | #637 |
| 11 | Command-R+ | 21 | #869 |
| 12 | Llama 4 Maverick Instruct | 20.89 | #467 |
| 13 | Kimi K2 Instruct (0905) | 20.89 | #267 |
| 14 | GPT-OSS-120B (Reasoning) | 20.67 | #346 (GPT-OSS-120B) |
| 15 | Yehia-7B-preview | 20.33 | #659 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=balsam-factuality · How It Works · Data refreshed daily, snapshot 2026-11-05.