SAHARA - Text Generation Cluster: leaderboard
Text-generation cluster score from the SAHARA African-language LLM leaderboard (UBC-NLP, ACL 2025): mean of spBLEU-1K on three machine-translation directions, paraphrase and title generation plus ROUGE-L on summarization. Scale 0-100, higher is better; absolute values are low.
Metric: Cluster score (mean of task scores). Source: huggingface.co. Saturation forecast: Around 2036. 48 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 18.24 | #71 |
| 2 | Gemini 2.5 Pro | 15.66 | #159 |
| 3 | GPT-5.6 Luna | 13.54 | #79 |
| 4 | Qwen 3.6 35B A3B | 12.42 | #238 |
| 5 | Claude Sonnet 4 | 12.04 | #207 |
| 6 | Qwen 3.6 27B | 11.96 | #182 |
| 7 | Qwen 3.5 27B | 11.73 | #192 |
| 8 | Llama 4 Scout Instruct | 11.45 | #547 |
| 9 | Gemini 2.5 Flash | 11.18 | #255 |
| 10 | Llama 3.1 70B Instruct | 11.15 | #567 |
| 11 | GPT-5 | 11.1 | #105 |
| 12 | Qwen 3.5 4B | 11.08 | #514 |
| 13 | GPT-4.1 | 11.06 | #258 |
| 14 | DeepSeek R1 Distill Llama 70B | 10.91 | #614 |
| 15 | Command-R+ (08-2024) | 10.85 | #838 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=sahara-text-generation-cluster · How It Works · Data refreshed daily, snapshot 2026-11-05.