Phi-4-reasoning-plus (Thinking): 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. The plus release adds reinforcement learning and retains its separately evaluated thinking configuration. Provider: Microsoft. Released 2025-04-30. Access: Open.
Aggregate Rating 1233 ± 15, rank #1967 of 2145 rated models, from 291 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AgentDrive - Comparative | 95 | Accuracy (%) | 97 |
| EuroEval Polish Summarization - PSC | 25.2 | chrF3++ (%) | 93.2 |
| AgentDrive - Scenario | 97.5 | Accuracy (%) | 93 |
| AgentDrive | 75.5 | Accuracy (%) | 86 |
| AI Energy Score (Text Generation) | 5 | Energy Score (1-5) | 79.5 |
| EuroEval Portuguese NLU - ScaLA PT | 24.82 | Matthews correlation coefficient (%) | 79.3 |
| AgentDrive - Policy | 92.5 | Accuracy (%) | 79 |
| Open Portuguese LLM - ASSIN2 RTE | 92.12 | Macro F1 (%) | 74.7 |
| SLMJury | 88.75 | Judge accuracy (%) | 73.3 |
| EuroEval Spanish NLU - ScaLA ES | 26.92 | Matthews correlation coefficient (%) | 68.1 |
| AgentDrive - Physics | 47.5 | Accuracy (%) | 68 |
| EuroEval Catalan NLU - ScaLA CA | 20.81 | Matthews correlation coefficient (%) | 66.4 |
Interactive version: theaggregate.ai/model?slug=phi-4-reasoning-plus-thinking · How It Works · Data refreshed daily, snapshot 2026-10-16.