@heyshrutimishra: Your app doesn’t need a cloud bill for every image. If every upload calls a paid vision API, usage becomes a running co…

Your app doesn’t need a cloud bill for every image.

If every upload calls a paid vision API, usage becomes a running cost.

MiniCPM-V 4.7 offers an open-source route to image understanding on your own hardware.

Here’s how I’d use it in an agent workflow. ↓

2/ Give your agent a folder of product photos.

The workflow I’d build:

Photo → MiniCPM-V 4.7 extracts visible attributes → the agent saves a structured catalogue entry.

Category, colour, pattern, readable label text.

An unclear detail stays blank for review. Start by replacing that one image-reading step in your app.

3/ Two more workflows worth testing:

• Expense app: receipt photo → merchant, date, line items and total → draft expense record.

• Screenshot library: image → visible text and topic tags → searchable index.

4/ This already reaches beyond desktop GPUs.

MiniCPM-V 4.7 comes in two sizes: the 1B used above, and 35B-A3B, the series’ first MoE model.

On an iPhone 17 Pro Max, the 35B-A3B card reports 13+ tokens/s using just 3 GB of memory.

On an iPad Pro 11“ M4 (16 GB), Lookahead Expert Scheduling takes it from 2.97 to 14.92 tokens/s.

The 1B’s 4-bit GGUF is about 672 MB, and both run locally on Ollama and llama.cpp.

These are the team’s own figures. Neither one comes with full test conditions.

5/ Building an app that processes images?

Pick one repeated task. Test accuracy and speed on your actual hardware, then compare the total cost with your current API.

Code and setup: https://github.com/OpenBMB/MiniCPM-V…

Every 4.7 model, including quantised builds: https://huggingface.co/collections/openbmb/minicpm-v-47…

How 35B-A3B runs on iPhone and iPad: https://github.com/tc-mb/llama.cpp-omni/blob/master/docs/les/lookahead-expert-scheduling.md…