Phi-3.5-vision-instruct — ONNX (INT4)

INT4-quantized ONNX export of Phi-3.5-vision-instruct, a 4.2B-parameter multimodal vision-language model from Microsoft. Accepts images and text prompts, generates text output. Optimized for CPU inference with int4 RTN block-32 quantization.

Mirrored for use with inference4j, an inference-only AI library for Java.

Original Source

• Repository: Microsoft
• License: mit

Usage with inference4j

try (VisionLanguageModel vision = VisionLanguageModel.builder()
        .model(ModelSources.phi3Vision())
        .build()) {
    GenerationResult result = vision.describe(Path.of("photo.jpg"));
    System.out.println(result.text());
}

Model Details

Property	Value
Architecture	Phi-3.5 Vision (4.2B parameters — CLIP ViT encoder + MLP projector + Phi-3 decoder)
Task	Image description, visual Q&A, multimodal chat
Context length	128K tokens
Quantization	INT4 RTN block-32 acc-level-4
ONNX files	3 models (vision encoder, embedding projector, text decoder)
Original framework	PyTorch (transformers)

License

This model is licensed under the MIT License. Original model by Microsoft.