Microsoft

Phi-4-multimodal-instruct

Multimodal
Zero-eval
#1ScienceQA Visual
#1BLINK
#1InterGPS
+2 more

by Microsoft

+
+
+
+
About

Phi-4 Multimodal was created to handle multiple input modalities including text, images, and potentially other formats. Built to extend Phi-4's efficiency into multimodal applications, it demonstrates that compact models can successfully integrate diverse information types.

+
+
+
+
Pricing Range
Input (per 1M)$0.05 -$0.05
Output (per 1M)$0.10 -$0.10
Providers1
+
+
+
+
Timeline
AnnouncedFeb 1, 2025
ReleasedFeb 1, 2025
Knowledge CutoffJun 1, 2024
+
+
+
+
Specifications
Training Tokens5.0T
Capabilities
Multimodal
+
+
+
+
License & Family
License
MIT
Performance Overview
Performance metrics and category breakdown

Overall Performance

15 benchmarks
Average Score
72.0%
Best Score
97.5%
High Performers (80%+)
7

Performance Metrics

Max Context Window
256.0K
Avg Throughput
25.0 tok/s
Avg Latency
1ms
+
+
+
+
All Benchmark Results for Phi-4-multimodal-instruct
Complete list of benchmark scores with detailed information
ScienceQA Visual
multimodal
0.97
97.5%
Self-reported
DocVQA
multimodal
0.93
93.2%
Self-reported
MMBench
multimodal
0.87
86.7%
Self-reported
POPE
multimodal
0.86
85.6%
Self-reported
OCRBench
multimodal
0.84
84.4%
Self-reported
AI2D
multimodal
0.82
82.3%
Self-reported
ChartQA
multimodal
0.81
81.4%
Self-reported
TextVQA
multimodal
0.76
75.6%
Self-reported
InfoVQA
multimodal
0.73
72.7%
Self-reported
MathVista
multimodal
0.62
62.4%
Self-reported
Showing 1 to 10 of 15 benchmarks
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+