Microsoft·released 2024-04-231 source
phi-3-medium 14B benchmark scores: 10 benchmarks tracked. 20% of its scores are independently reproduced — the rest are self-reported or unverified.
Leads on 0 of 10 benchmarks·best result 88.8 on ARC AI2 (self-reported)·2 of 10 independently reproduced
| Benchmark | Score | Evidence status | Measured |
|---|---|---|---|
| ARC AI2 | 88.8accuracy (%) | self-reported· optimizedT1 | 2024-04-23Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| OpenBookQA | 83.2accuracy (%) | self-reported· optimizedT1 | 2024-04-23Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| HellaSwag | 76.53accuracy (%) | self-reported· optimizedT1 | 2024-04-23Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| BBH | 75.2accuracy (%) | self-reported· optimizedT1 | 2024-04-23Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| TriviaQA | 73.9accuracy (%) | self-reported· optimizedT1 | 2024-04-23Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| MMLU | 70.67accuracy (%) | unverified· optimizedT1 | 2024-04-23Stanford CRFM Leaderboard |
| Winogrande | 63accuracy (%) | self-reported· optimizedT1 | 2024-04-23Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| ANLI | 33.7accuracy (%) | self-reported· optimizedT1 | 2024-04-23Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone |
| MATH level 5 | 17.56accuracy (%) | reproduced· optimizedT1 | 2024-04-23Epoch AI |
| GPQA diamond | 3.45accuracy (%) | reproduced· optimizedT1 | 2024-04-23Epoch AI |
Claims drawn from cited facts, not live model generation.
Developed by Microsoft, this model is a mid-size 14-billion-parameter model released in 2024. That mid-size scale provides more headroom than compact alternatives while remaining more efficient and less costly than large frontier-scale models.
4 cited facts
This model is scored on 10 tracked benchmarks, but it holds no current top score on any of them. The recorded results are independently reproduced, meaning outside evaluation backs the scores up and the record can be trusted as externally verified.
3 cited facts
Its mean gap to the best recorded score across its tracked benchmarks is 25.17 points.
1 cited fact
2 facts cross-checked across data sources: 2 single-source.
Source facts, citations, and refresh stamp for this record.
Sources: curated_param_counts, epoch_benchmarks
phi-3-medium 14B is an AI model developed by Microsoft, released 2024-04-23. Its benchmark record below tags every score as independently reproduced, self-reported, or unverified.
phi-3-medium 14B has recorded scores on 10 benchmarks, each shown with its evidence status.
phi-3-medium 14B has recorded scores on 10 benchmarks — OpenBookQA, Winogrande, GPQA diamond, MATH level 5, TriviaQA, MMLU, and 4 more. The full table above shows each score with its evidence status.
2 of 10 recorded scores (20%) are independently reproduced rather than self-reported by the lab.
phi-3-medium 14B has 10 tracked claims: 2 independently reproduced, 7 self-reported, 1 unverified.