Phi-4-mini (Reasoning): benchmark results
Microsoft's Phi-4-mini-reasoning, a lightweight model in the Phi-4 family fine-tuned on reasoning-dense synthetic data for multi-step mathematical problem-solving, supporting a 128K-token context. Provider: Microsoft. Released 2025-04-30. Access: Open.
Aggregate Rating 1206 ± 25, rank #2032 of 2168 rated models, from 118 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| HarmActionsEval | 2.84 | SafeActions@1 (imported) | 72.2 |
| EuroEval English NLU - CoNLL EN | 69.04 | Micro F1 excluding MISC (%) | 50.6 |
| Syslog Severity Classification - Zero-Shot | 9.2 | Accuracy on 9,355 held-out journalctl logs, exact syslog sev | 50 |
| EuroEval Portuguese Knowledge | 39.64 | Knowledge Average Score (%) | 48 |
| EuroEval French Knowledge | 41.54 | Knowledge Average Score (%) | 43.7 |
| EuroEval German Knowledge | 38.36 | Knowledge Average Score (%) | 41.8 |
| EuroEval Spanish Knowledge | 35.22 | Knowledge Average Score (%) | 41.1 |
| ZeroEval MATH-500 | 94.6 | MATH-500 Score | 40.6 |
| EuroEval Swedish Knowledge | 30.63 | Knowledge Average Score (%) | 40.4 |
| EuroEval Spanish NLU - ScaLA ES | 5.8 | Matthews correlation coefficient (%) | 39.9 |
| EuroEval Dutch Knowledge | 35.12 | Knowledge Average Score (%) | 39.2 |
| EuroEval Italian Knowledge | 32.58 | Knowledge Average Score (%) | 39.1 |
Interactive version: theaggregate.ai/model?slug=phi-4-mini-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-20.