Is there a public workflow to convert Hugging Face models into Nexa `.nexa` NPU model packages?

[vjeson]

Hi NexaSDK team,

I am currently testing the Android NexaSDK OCR example on a Qualcomm Snapdragon device.

The current OCR model I am using is:

• NexaAI/paddleocr-npu-mobile
• Android SDK dependency: ai.nexa:core
• Runtime/plugin: plugin_id = "npu"
• Model files include:
  • weights-1-1.nexa
  • attachments-1-1.nexa
  • nexa.manifest
  • ppocr_keys_v1.txt

This works well on device. However, the model card says it is based on PaddleOCR v4 / PP-OCRv4.

PaddleOCR has now released PP-OCRv6, and the latest models are available on Hugging Face, for example:

• https://huggingface.co/collections/PaddlePaddle/pp-ocrv6
• ONNX / Paddle / safetensors model variants are available there.

My questions are:

1. Is there any public tool or documented workflow to convert a Hugging Face model, such as an ONNX or Paddle model, into Nexa’s .nexa model package format?
2. For NPU models using plugin_id = "npu", is the .nexa package generated only by the Nexa/Qualcomm team, or can developers generate it themselves?
3. If self-conversion is supported, what is the recommended path?
   • Nexa CLI?
   • Qualcomm AI Hub / AI Hub Workbench?
   • QNN / QAIRT tools?
   • Another official packaging tool?
4. If Qualcomm AI Hub can compile a model into QNN DLC, QNN context binary, or precompiled QNN ONNX, can those artifacts be used directly by NexaSDK’s Android CvWrapper, or is an additional Nexa-specific packaging step required?
5. For upgrading NexaAI/paddleocr-npu-mobile from PP-OCRv4 to PP-OCRv6, should developers wait for an official Nexa/Qualcomm converted model package, or is there a supported way to perform the conversion independently?

The documentation currently explains how to load existing NPU models, but I could not find a public guide for creating .nexa packages from new models.

Any clarification on the official model conversion and packaging workflow would be very helpful.

Thanks!