Fähigkeiten großer Sprachmodelle (Beispieldaten).
Composite scores may decompose into reasoning, coding, multilingual, and safety dimensions; Methodology must cite benchmark versions and tool-use policy.
Public ranking policy: rows are sorted by composite score (desc). Composite score is a weighted sum of normalized sub-metrics; ties are broken by higher recent activity.
| Rang | Modell | Anbieter | Größe | Punktzahl | Hinweise |
|---|---|---|---|---|---|
| 1 | Nova-Large-2 | Nova AI | ~400B MoE | 95 | Reasoning-Modus |
| 2 | Summit-Pro | Summit | ~200B | 93.4 | Starke Befolgung von Anweisungen |
| 3 | DeepLine-R1 | DeepLine | ~70B | 91.9 | Offene Gewichte |
| 4 | Cedar-32B | Cedar | 32B | 89.7 | Ausgewogen ZH/EN |
| 5 | Birch-Mini | Birch | 8B | 87.3 | On-Device-Deployment |
| 6 | Fjord-1.5 | Fjord Labs | 14B | 86.1 | Tool-Aufrufe |
| 7 | Ridge-Code | Ridge | 33B | 85 | Code-fokussiert |
| 8 | Willow-Base | Willow | 3B | 82.4 | Sehr geringe Latenz |