Atlas

Models

← All models

Phi-4-multimodal-instruct

Microsoft Research · 2025-02-26 · 5.6B parameters

Phi-4-multimodal-instruct is Microsoft's open-weight, instruction-tuned 5.6B-class multimodal model, publicly launched on February 26, 2025. It shares Phi-4-Mini-Instruct's 32-layer GQA language backbone and adds image and audio encoders, projectors, and modality-specific LoRA adapters with routing for text-only, vision-language, speech/audio, and combined vision-speech inputs, while generating text. The checkpoint supports a 131,072-token context with LongRoPE and is intended for edge and resource-constrained deployments.

Benchmark scores

BenchmarkScore
Artificial Analysis Intelligence Index4.5
Artificial Analysis Openness Index50.0
Audio MultiChallenge (Text Output)15.5
Capability108.5
GPQA Diamond31.5
Humanity's Last Exam (Text-Only)4.4
LiveCodeBench (Artificial Analysis)13.1
Longform Creative Writing22.8
MATH-500 (Artificial Analysis source)69.3
MathVista (mini)62.4
MMLU Pro48.5
MMMU Pro14.5
SciCode (Artificial Analysis)11.0
SnakeBench Average Apples0.9
SnakeBench Best Apples2.0
SnakeBench Rating20.3
SnakeBench Total Apples24.0
SnakeBench Win Rate40.7
Loading Atlas data…