I tried multiple times to export the "Phi-4-multimodal-instruct" model to OpenVINO IR without success. My goal is to use "optimum-intel" to quantize the model in order to be able to run it on Intel NPU. According to this OpenVINO Documentation page, I should be able to use the "Phi-4-multimodal-instruct" on my Intel Core Ultra NPU.

I tried following the official instructions to prepare models for NPU Execution: https://docs.openvino.ai/2025/openvino-workflow-generative/inference-with-genai/inference-with-genai-on-npu.html and they didn't work. Of course, the INT4-CW quantization method is being used.

I have tried a clean Python 3.12 with "optimum-intel" installed following instructions on PyPi page as well as the OpenVINO phi-4-multimodal noteook but both couldn't produce a valid output.

I am now using a custom environment which allowed to produce a valid output using "image-text-to-text", yet the "automatic-speech-recognition" task still has issues. Here is the custom environment packages I'm using with a Python 3.12 venv:

requirements.txt

Downloading the model first successfully allowed me to convert it, specifying the task "image-text-to-text": optimum-cli export openvino -m Phi-4-multimodal-instruct --weight-format int4 --sym --ratio 1.0 --group-size -1 --trust-remote-code --task image-text-to-text Phi-4-multimodal-instruct. Because I was using transofmers>=4.50 I had to apply a patch as officially specified here. This output worked fine with VLMPipeline and LLMPipeline classes of OpenVINO GenAI.

Here comes the issue

According to optimum-intel, the "automatic-speech-recognition" task should be supported too for this model (ValueError: Asked to export a phi4mm model for the task text-to-text, but the Optimum OpenVINO exporter only supports the tasks image-text-to-text, automatic-speech-recognition for phi4mm. Please use a supported task) but I couldn't manage to get it working. By default the inferred task was "automatic-speech-recognition" and always ended with the following error: This means that exporting the model with Optimum without any parameter would also result in this error:

RuntimeError: Exception from src\inference\src\cpp\core.cpp:97: Check 'util::directory_exists(path) || util::file_exists(path)' failed at src\frontends\common\src\frontend.cpp:117: FrontEnd API failed with GeneralFailure: ir: Could not open the file: "C:\Users\<username>\AppData\Local\Temp\tmp7p952plq\openvino_encoder_model.xml"

I also tried removing all the NPU-specific parameters leaving only OpenVINO export config but the result was the same, therefore I decided to open this issue.