Microsoft Research
Phi-4 benchmark results
9 results on 8 benchmarks from 5 sources.
Also listed as: phi-4 · Phi-4 14b · Phi-4 (Prompt) · Phi-4-reasoning · microsoft/phi-4
All benchmark results
| Benchmarks | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|
| LMArena Text | 1256 | 318 | Overall, style control on; 24126 votes; 95% CI 1252–1261“phi-4” | Official leaderboard | Oct 8, 2026 |
| Open LLM Leaderboard v2 | 30.4% | 969 | Average of 6 normalised benchmarks; bfloat16; 14.66B params“microsoft/phi-4” | Official leaderboard | Mar 13, 2025 |
| Open LLM Leaderboard v2 | 29.5% | 1112 | Average of 6 normalised benchmarks; float16; 14.66B params“microsoft/phi-4” | Official leaderboard | Mar 13, 2025 |
| GPQA Diamond | 56.1% | 190 | Epoch AI run, accuracy | Independent run | Jan 31, 2025 |
| MMLU | 84.8% | 7 | Exact match; 5-shot | Paper / tech report via Epoch AI | Dec 12, 2024 |
| MMLU-Pro | 70.4% | 99 | Overall accuracy; data source: Self-Reported; 14B | Vendor-reported | Mar 11, 2026 |
| AIME (OTIS Mock AIME 2024–2025) | 13.8% | 213 | Epoch AI run, OTIS Mock AIME 2024–2025 | Independent run | Feb 25, 2025 |
| MATH Level 5 | 64.9% | 47 | Epoch AI run, accuracy | Independent run | Jan 31, 2025 |
| BFCL v4 (function calling) | 28.8% | 70 | Overall accuracy; prompt-based“Phi-4 (Prompt)” | Official leaderboard | Oct 17, 2026 |
Related on this site
FAQ
Where do these Phi-4 scores come from?
From Epoch AI Benchmarking Hub, Berkeley Function Calling Leaderboard (BFCL), LMArena leaderboard dataset, MMLU-Pro leaderboard (TIGER-Lab), and Hugging Face Open LLM Leaderboard v2 (archived). Each row links to the original leaderboard, dataset or report, and vendor-reported numbers are labelled.