Trinity-Mini-Base: benchmark results
Arcee AI Trinity Mini base checkpoint before fine-tuning: 32 layers, two dense and 30 MoE layers, routing to eight of 128 experts plus one shared expert. Its publisher reports 26B total, 3B active parameters and 10T training tokens. Provider: Arcee AI. Released 2025-12-01. Access: Open.
Aggregate Rating 1303 ± 29, rank #8677 of 10916 rated models, from 203 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EuroEval Dutch NLU - CoNLL NL | 62.38 | Micro F1 excluding MISC (%) | 69.2 |
| EuroEval Portuguese NLU - HAREM | 45.78 | Micro F1 excluding MISC (%) | 65.4 |
| EuroEval Ukrainian NLU - NER UK | 52.99 | Micro F1 excluding MISC (%) | 62 |
| EuroEval Finnish NLU - Turku NER FI | 54.7 | Micro F1 excluding MISC (%) | 61.9 |
| EuroEval Czech NLU - PONER | 49.41 | Micro F1 excluding MISC (%) | 61.8 |
| EuroEval Estonian NLU - EstNER | 47.02 | Micro F1 excluding MISC (%) | 61.2 |
| EuroEval Slovak NLU - UNER SK | 52.82 | Micro F1 excluding MISC (%) | 59.6 |
| EuroEval Latvian NLU - Fullstack NER LV | 45.86 | Micro F1 excluding MISC (%) | 59.5 |
| EuroEval Icelandic NLU - MIM-GOLD NER | 47.88 | Micro F1 excluding MISC (%) | 52.4 |
| EuroEval Slovene NLU - SSJ500K NER | 50.38 | Micro F1 excluding MISC (%) | 52.2 |
| EuroEval Swedish NLU - SUC3 | 55.56 | Micro F1 excluding MISC (%) | 51.8 |
| EuroEval German NLU - GermEval | 57.63 | Micro F1 excluding MISC (%) | 50.5 |
Interactive version: theaggregate.ai/model?slug=trinity-mini-base · How It Works · Data refreshed daily, snapshot 2026-11-08.