Phi-3 Mini 128K Instruct: benchmark results
Microsoft's Phi-3-Mini-128K-Instruct, a 3.8B model trained on synthetic and filtered web data, post-trained with SFT and DPO; the 128K-context variant of Phi-3 Mini (4K also available). Provider: Microsoft. Released 2024-04-23. Access: Open.
Aggregate Rating 1379 ± 12, rank #1298 of 1656 rated models, from 149 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard v1 - MMLU | 68.7 | Accuracy (%) (5-shot) | 91.8 |
| Cyber Bench Hub (v2) | 98 | Accuracy (%) | 90.5 |
| Open LLM Leaderboard v1 - GSM8K | 69.52 | Accuracy (%) (5-shot) | 89.5 |
| MERA - ruHumanEval | 38.35 | pass@1 (%) | 79.9 |
| Open LLM Leaderboard - BBH | 55.75 | Score | 76.4 |
| MERA - ruCodeEval | 36.28 | pass@1 (%) | 76.2 |
| BiGGen-Bench | 3.68 | Average Score (1-5) | 75.5 |
| Enkrypt AI - Jailbreak Risk | 5.53 | Risk Score | 74.3 |
| Open LLM Leaderboard - GPQA | 31.8 | Score | 73.8 |
| Open LLM Leaderboard - IFEval | 59.76 | Score | 71.7 |
| Enkrypt AI - Safety Risk | 23.89 | Risk Score | 63.6 |
| BABILong (NIAH, v0, 8k) | 50 | QA1–5 Mean Accuracy (%) | 62.5 |
Interactive version: theaggregate.ai/model?slug=phi-3-mini-128k-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.