Phi-4 Mini Instruct: benchmark results

Microsoft's MIT-licensed 3.8B dense instruct model in the Phi-4 family (February 2025), with 128K context, grouped-query attention, and a 200K-token vocabulary. Provider: Microsoft. Released 2025-02-26. Access: Open.

Aggregate Rating 1394 ± 8, rank #1245 of 1656 rated models, from 678 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
Open Arabic LLM - Aratrust Unfairness94.55Accuracy (%)91
Open LLM Leaderboard - IFEval73.78Score88.4
LatamBoard - Spanish MGSM14.4Score (%)87.9
Grip on LLMs - HonestCityBench - Outdated Information52Appropriate acknowledgement rate (%; LLM judge)86.7
Grip on LLMs - Dutch Simplification - Amsterdam42.48SARI (0-100)83.3
Grip on LLMs - HonestCityBench - Non-Text Modality37Appropriate acknowledgement rate (%; LLM judge)83.3
Holtercare-Bench - Heart Rate Extremum Timing (Text)36.31Accuracy (%; multiple-choice closed QA on the millisecond ti83.3
Open Arabic LLM - Alghafa Multiple Choice Rating Sentiment NO Neutral Task80.1Accuracy (%)83.3
Grip on LLMs - Dutch Simplification Average40.99SARI (0-100; mean of two datasets)80
Enkrypt AI - Safety Risk21.33Risk Score79.3
Open Arabic LLM - Alghafa Multiple Choice Grounded Statement Xglue Mlqa Task90.67Accuracy (%)78.7
Grip on LLMs - HonestCityBench - Average33Appropriate acknowledgement rate (%; mean of five categories78.3

Interactive version: theaggregate.ai/model?slug=phi-4-mini-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.