Phi-3 Mini Instruct 3.8B: benchmark results
Microsoft's Phi-3-Mini-Instruct, a 3.8B dense decoder-only model trained on the Phi-3 synthetic-plus-filtered-web dataset, post-trained with SFT and DPO; released in 4K- and 128K-context variants. Provider: Microsoft. Released 2024-04-23.
Aggregate Rating 1388 ± 41, rank #6611 of 10916 rated models, from 9 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AA Humanity's Last Exam | 4.96 | Accuracy (%) | 26.3 |
| AA MMLU-Pro | 43.51 | Accuracy (%) | 10.6 |
| AA LiveCodeBench | 11.64 | Pass@1 (%) | 10.5 |
| AA GPQA Diamond | 31.92 | Accuracy (%) | 8.1 |
| AA IFBench | 23.88 | Accuracy (%) | 6.2 |
| AA Terminal-Bench Hard | 0 | Accuracy (%) | 5.5 |
| Tau2-Bench Telecom | 0 | Success Rate (%) | 2.6 |
| AA TAU-2 Bench | 0 | Accuracy (%) | 2.5 |
| AA AIME 2025 | 0.33 | Accuracy (%) | 1.7 |
Interactive version: theaggregate.ai/model?slug=phi-3-mini-instruct-3-8b · How It Works · Data refreshed daily, snapshot 2026-11-08.