AFM-4.5B-Base: benchmark results
Arcee AI AFM base checkpoint after merging and context extension, with a 36-layer dense decoder using grouped-query attention and ReLU-squared activation. Its card describes 8 trillion pretraining and midtraining tokens, split into 6.5T general and 1.5T math/code-focused tokens. Provider: Arcee AI. Released 2025-09-01. Access: Open.
Aggregate Rating 1352 ± 29, rank #7643 of 10916 rated models, from 240 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EuroEval German NLU - Germanquad | 57.18 | Character F1 (%) | 73.8 |
| EuroEval French NLU - FQuAD | 69.43 | Character F1 (%) | 72.4 |
| EuroEval English NLU - SST-5 | 67.01 | Matthews correlation coefficient (%) | 69.2 |
| EuroEval English NLU - ScaLA EN | 49.54 | Matthews correlation coefficient (%) | 68.8 |
| EuroEval Polish NLU - Polemo2 | 87.19 | Matthews correlation coefficient (%) | 67 |
| EuroEval Polish NLU - PoQuAD | 48.44 | Character F1 (%) | 66.3 |
| EuroEval English Knowledge | 80.5 | Knowledge Average Score (%) | 64.8 |
| EuroEval Spanish NLU - MLQA ES | 61.63 | Character F1 (%) | 61.1 |
| EuroEval Catalan NLU - MultiWikiQA CA | 64.42 | Character F1 (%) | 55.5 |
| EuroEval Italian NLU - SQuAD IT | 67.44 | Character F1 (%) | 54.6 |
| EuroEval Dutch NLU - SQuAD NL | 73.23 | Character F1 (%) | 54.3 |
| EuroEval Portuguese NLU - MultiWikiQA PT | 69.46 | Character F1 (%) | 53.9 |
Interactive version: theaggregate.ai/model?slug=afm-4-5b-base · How It Works · Data refreshed daily, snapshot 2026-11-08.