MiniCPM-V 4.7 1B photo questions on iPhone — integration demo

[1031242365]

I’m building Local LLM, an iPhone app for running models on-device, and wanted to share MiniCPM-V 4.7 1B running in the app.

I gave it a photo of building blocks and asked for three things to build. The clip below puts its answer next to Bonsai 2 4B on the same photo: Bonsai suggests a tower, a vehicle and a miniature town; MiniCPM suggests a house, a toy shelf and a figure or animal.

This is a small everyday-use example, not a benchmark. The prompts differ slightly, and input/loading/generation waits have been cut, so the side-by-side timing is not a speed comparison.

Standalone MiniCPM recording: https://x.com/TonyZ278264/status/2111614812790681938

App: https://apps.apple.com/app/id6804004948

Thanks to the MiniCPM team for making small on-device vision models available. What everyday photo tasks would be useful to try next?

bonsai-minicpm-comparison-v2.mp4

[tc-mb]

I’m very pleased to receive this application overview from the community.
Let’s keep this discussion thread open, so that community members can find tasks that interest them.

[1031242365]

Thanks! It’s been fun trying MiniCPM-V on my iPhone. I’ll keep sharing examples here as I test more photo questions in Local LLM.

Are there any particular tasks you’d like to see tried on a phone? Happy to record a few more.