Phi-4 (Reasoning): benchmark results
Microsoft Research's Phi-4 reasoning derivative, post-trained on chain-of-thought traces focused on math, science and coding. It remains distinct from the original Phi-4 checkpoint. Provider: Microsoft. Released 2025-04-30. Access: Open.
Aggregate Rating 1417 ± 18, rank #1368 of 2168 rated models, from 155 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EuroEval Icelandic NLU - NQII | 55.92 | Character F1 (%) | 87.2 |
| EuroEval Finnish NLU - Tydiqa FI | 70.57 | Character F1 (%) | 85 |
| EuroEval German NLU - Germanquad | 58.9 | Character F1 (%) | 78.1 |
| EuroEval Portuguese NLU - MultiWikiQA PT | 73.65 | Character F1 (%) | 77.1 |
| MIST (Selective Trust) | 88.3 | Overall Accuracy (%) | 68.2 |
| Open Portuguese LLM - FaQuAD NLI | 73.17 | Macro F1 (%) | 68.1 |
| Open Portuguese LLM - BLUEX | 56.88 | Accuracy (%) | 67.6 |
| Medmarks - MedHallu Hard | 48.36 | Score (%) | 67 |
| SLMJury | 88.24 | Judge accuracy (%) | 66.7 |
| Open Portuguese LLM - ASSIN2 RTE | 91.32 | Macro F1 (%) | 66.3 |
| EuroEval Italian NLU - SQuAD IT | 68.84 | Character F1 (%) | 61.8 |
| Medmarks - LongHealth Task 2 | 88 | Score (%) | 60.8 |
Interactive version: theaggregate.ai/model?slug=phi-4-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-20.