Add Moondream2 vision-language notebook with OpenVINO

[LuciferDono]

Summary

• Add new notebook demonstrating Moondream2 (2B parameter vision-language model) with OpenVINO
• Model conversion to OpenVINO IR with INT4 weight compression via optimum-cli
• Inference demos: image captioning, visual question answering, object detection
• Interactive Gradio interface for uploading images and asking questions
• Follows existing notebook conventions (SmolVLM2, Phi-4 multimodal patterns)

Details

Moondream2 is a compact VLM designed for edge deployment. This notebook shows how to optimize it with OpenVINO for efficient inference on Intel hardware. Uses OVModelForVisualCausalLM from optimum-intel, device_widget for device selection, and INT4/FP16 compression options.

Checklist

• Notebook runs end-to-end
• Follows repository conventions (telemetry, utils fetch, device widget, Gradio demo)
• README with Scarf pixel
• Cross-platform compatible (pathlib.Path, no hardcoded paths)
• Cleanup cell included

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[LuciferDono]

hey @aleksandr-mokrov, just checking in on this. anything I should change or is there something blocking review? happy to adjust if needed.

[aleksandr-mokrov]

hey @aleksandr-mokrov, just checking in on this. anything I should change or is there something blocking review? happy to adjust if needed.

Thanks for the contribution!

I tried running the notebook locally and hit an error during model conversion with optimum-cli. Could you please share the full conversion log from a successful run on your side? Also, which versions of optimum-intel and openvino did you use?

[LuciferDono]

Hey @aleksandr-mokrov, sorry for the wait.

Turns out Moondream2 registers as moondream1, which isn't in optimum-intel's export registry. The trust_remote_code error was actually from a positional arg bug, but even with that fixed the real problem is that the architecture just isn't supported by optimum-cli.

Looking into converting it with ov.convert_model directly instead. Will update the PR when I have something working.