Phi-4 Multimodal Instruct: benchmark results
Microsoft's Phi-4-multimodal-instruct, a lightweight multimodal model processing text, image and audio inputs to text output, with 128K context and SFT, DPO and RLHF alignment. Provider: Microsoft. Released 2025-02-26. Access: Open.
Aggregate Rating 1466 ± 6, rank #906 of 1656 rated models, from 580 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AVI-Bench - Audio-Referenced Visual Hallucination | 86.8 | Task score (%) on audio-referenced visual hallucination (250 | 100 |
| AppTek Call-Center Dialogues - Indian English | 9.4 | Word error rate (%) on the Indian-accented English (en_IN) c | 100 |
| AudioBench - AISHELL ASR ZH WER | 7.47 | WER (%) (lower is better) | 100 |
| AudioBench - CoVoST2 EN-ZH BLEU | 45.3 | Mean Score (%) | 100 |
| AudioBench - CoVoST2 ZH-EN BLEU | 22.68 | Mean Score (%) | 100 |
| AudioBench - LibriSpeech Clean WER | 1.68 | WER (%) (lower is better) | 100 |
| AudioBench - MMAU Music | 68.86 | Score (%) | 100 |
| AudioBench - TED-LIUM 3 WER | 2.85 | WER (%) (lower is better) | 100 |
| MUN-lang (5-Shot) | 65.1 | Win rate (%; share of MUN-lang items where a GPT-4o judge, A | 100 |
| MUN-vis (5-Shot) | 63 | Win rate (%; share of MUN-vis items where a GPT-4o judge, Al | 100 |
| MUN-vis (Zero-Shot) | 41 | Win rate (%; share of MUN-vis items where a GPT-4o judge, Al | 100 |
| MultivationBench | 39.22 | Exact-match accuracy (%, predicted option set equals the gol | 100 |
Interactive version: theaggregate.ai/model?slug=phi-4-multimodal-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.