Phi-4-reasoning places in the top three on HumanEval+.
Microsoft Apr 2025 Open weights Small
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Arena-Hard-Auto | 73.3% | #24 of 39 | Lab model card | |
| MMLU-Pro | 74.3% | #111 of 167 | Lab model card | |
| IFEval | 83.4% | #93 of 135 | Lab model card | |
| FlenQA | 97.7% | #2 of 2 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| 3SAT-Search | 78% | #1 of 2 | Lab model card | |
| BA-Calendar | 67.7% | #1 of 2 | Lab model card | |
| AIME 2024 | 75.3% | #53 of 88 | Lab model card | |
| GPQA Diamond | 67.1% | #193 of 302 | Lab model card | |
| Omni-MATH | 76.6% | #3 of 4 | Lab model card | |
| AIME 2025 | 63.1% | #122 of 143 | Lab model card | |
| HMMT 2025 | 43.8% | #52 of 54 | Lab model card | |
| AIME 1983-2024 | 83.1% | #2 of 2 | Lab model card | |
| TSP-Opt | 37.5% | #2 of 2 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| HumanEval+ | 92.9% | #3 of 27 | Lab model card | |
| Codeforces rating | 1,736 | #34 of 47 | Lab model card | |
| LiveCodeBench v5 | 53.8% | #36 of 47 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SpatialEval Maze-Nav | 55.1% | #1 of 2 | Lab model card | |
| SpatialEval Spatial-Map | 73.7% | #1 of 2 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ToxiGen (discriminative) | 85.5% | #1 of 2 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| PhiBench | 70.6% | #2 of 3 | Lab model card |
The Self-Reported Index fits one score to the 9 capability results Microsoft reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 127.7 to 131.9. How we compute it.
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
The benchmarks the lab chose to headline, with the scores it gave.
Every result in the longer documents the lab published with it. 9 of them aren't in the launch post.