New: connect Claude & other AIs to GenAIList over MCP — research the catalog and contribute to the shared knowledge base. Learn how →
A head-to-head benchmark comparison of Phi-2 across 10 evaluations.
Over the past few months, our Machine Learning Foundations team at Microsoft Research has released a suite of small language models (SLMs) called “Phi” that ach.
Full Phi-2 specs →| Benchmark | Phi-2 |
|---|---|
| code | |
| HumanEval | 47.6% |
| MBPP | 55% |
| general | |
| MMLU | 52.7% |
| knowledge | |
| ARC-Challenge | 61.1 |
| math | |
| GSM8K | 57.2% |
| MATH | 3.5% |
| reasoning | |
| BIG-Bench Hard | 43.4% |
| HellaSwag | 73.1 |
| Winogrande | 74.4 |
| safety | |
| TruthfulQA | 44.5% |
Best result per row highlighted in cyan. Each benchmark links to its definition and sources; each model links to its full scorecard.
Phi-2 has the strongest result on HumanEval among the models compared here. See the code-category rows in the table for the full picture.
This comparison covers 10 benchmarks on which at least one of the selected models has a published score.
Scores are aggregated from official model cards, technical reports and standard public evaluations, and link back to each benchmark's source. They are updated as new results are published.
Help us grow the database! Only the announcement link is required — we'll fetch the rest.
Missing a card? Drop the spec/announcement link — we'll verify the specs and add it. The link is the only required field.
Know an inference runtime we're missing (llama.cpp, vLLM, MLX, ComfyUI…)? Drop the link.
Spotted a wrong spec or an out-of-date price? Tell us and we'll verify it.
We'll send you a login code via email
