SAHARA - Token-Level Cluster: leaderboard
Token-level cluster score from the SAHARA African-language LLM leaderboard (UBC-NLP, ACL 2025): mean macro-F1 of named entity recognition, phrase chunking and part-of-speech tagging in African languages. Scale 0-100, higher is better.
Metric: Cluster score (mean of task scores). Source: huggingface.co. Saturation forecast: Around February 2027. 48 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 73.18 | #71 |
| 2 | Gemini 2.5 Pro | 65.99 | #159 |
| 3 | GPT-5 | 65.07 | #105 |
| 4 | GPT-5.6 Luna | 61.65 | #79 |
| 5 | GPT-5.1 (2025-11-13) | 61.59 | #107 |
| 6 | Claude Sonnet 4 (20250514) | 44.86 | #223 |
| 7 | Claude Sonnet 4 | 38.85 | #207 |
| 8 | Gemini 2.5 Flash | 34.91 | #255 |
| 9 | GPT-4.1 | 34.05 | #258 |
| 10 | Qwen 3.5 27B | 33.84 | #192 |
| 11 | Qwen 3.6 27B | 32.43 | #182 |
| 12 | Gemma 4 31B (IT) | 28.33 | #178 |
| 13 | c4ai-command-a-03-2025 | 25.16 | #468 |
| 14 | Qwen 3.6 35B A3B | 23.95 | #238 |
| 15 | Gemma 4 26B A4B (IT) | 20.89 | #293 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=sahara-token-level-cluster · How It Works · Data refreshed daily, snapshot 2026-11-05.