SAHARA - Paraphrase: leaderboard
Paraphrase generation in African languages, scored against reference paraphrases, from the SAHARA African-language LLM leaderboard (UBC-NLP, ACL 2025). The task score is the mean over the languages the task covers; spBLEU-1K on a 0-100 scale, higher is better.
Metric: spBLEU-1K. Source: huggingface.co. 48 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 4B 2507 Instruct | 36.29 | #768 |
| 2 | Command-R+ (08-2024) | 36.06 | #838 |
| 3 | Qwen 3.5 4B | 35.51 | #514 |
| 4 | Gemma 4 E2B (IT) | 34.9 | #779 |
| 5 | DeepSeek R1 Distill Qwen 32B | 32.13 | #662 |
| 6 | Qwen 3.5 2B | 32.08 | #1114 |
| 7 | aya-23-8B | 31.68 | #1267 |
| 8 | Qwen 3 30B A3B 2507 Instruct | 31.31 | #492 |
| 9 | Gemma 2 2B (IT) | 31.15 | #1328 |
| 10 | Gemma 4 31B (IT) | 31 | #178 |
| 11 | aya-23-35B | 30.68 | #873 |
| 12 | Qwen 3.6 35B A3B | 29.72 | #238 |
| 13 | Qwen 3.5 9B | 29.02 | #391 |
| 14 | Llama 3.2 3B Instruct | 28.88 | #1327 |
| 15 | Llama 3.1 8B Instruct | 28.73 | #1017 |
Interactive version: theaggregate.ai/benchmark?slug=sahara-paraphrase · How It Works · Data refreshed daily, snapshot 2026-11-05.