BALSAM - Summarization: leaderboard
BALSAM Arabic benchmark (SDAIA/KSAA, dataset version 2), Summarization category: producing a summary that keeps a text's key information and concepts. The site's overall score blends LLM-judged generated answers with multiple-choice answers across the category's tasks; 0-100, higher is better.
Metric: Overall score (0-100, LLM-judged generation and multiple choice). Source: benchmarks.ksaa.gov.sa. Saturation forecast: Around 2032. 37 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 235B A22B | 52.64 | #323 |
| 2 | GLM-4.7 (Reasoning) | 52.09 | #202 (GLM-4.7) |
| 3 | GLM-5 (Thinking) | 51.9 | #149 (GLM-5) |
| 4 | GPT-5.2 | 51.55 | #115 |
| 5 | Kimi K2 Instruct (0905) | 51.44 | #267 |
| 6 | DeepSeek V3.2 | 50.46 | #210 |
| 7 | AceGPT-v2-8B-Chat | 50.25 | #746 |
| 8 | GPT-OSS-120B (Reasoning) | 50.04 | #346 (GPT-OSS-120B) |
| 9 | Claude Opus 4.5 | 49.86 | #87 |
| 10 | Grok 4.1 Fast (Reasoning) | 49.71 | #227 (Grok 4.1 Fast) |
| 11 | Command A 03 2025 | 49.35 | #530 |
| 12 | Fanar-C-2-27B | 49.17 | #637 |
| 13 | Command-R+ | 48.77 | #869 |
| 14 | Llama 4 Maverick Instruct | 48.11 | #467 |
| 15 | Nova Pro (v1) | 47.88 | #508 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=balsam-summarization · How It Works · Data refreshed daily, snapshot 2026-11-05.