Player not loading? Watch on YouTube
This tutorial installs MiniCPM-V 4.6 locally and tests handwritten OCR, financial document extraction and video description. The presenter describes OpenBMB's 1.3 billion parameter model as suitable for on-device use on iOS, Android and HarmonyOS. The demonstration uses Ubuntu with an Nvidia RTX A6000 and 48 GB of VRAM; it does not test those mobile platforms.
The setup uses a virtual environment, prerequisite installation and a Jupyter notebook to download and load the model. Inference code largely follows the Hugging Face example, with cosmetic changes. The presenter reports an initial GPU memory footprint just above 1 GB and suggests CPU use with sufficient RAM, including 16 GB, but does not demonstrate CPU performance.
The OCR test exposes a practical limitation for local AI deployment: reported VRAM rises above 26 GB while the model processes a handwritten letter. The presenter praises the transcription's handling of words and punctuation, but notes the delay. He says image requests can take one to two minutes or longer, depending on the prompt and image.
The financial statement test asks for the total appropriation excluding special accounts for 2010-11. The presenter judges the extracted answer and reasoning correct. A final test describes an AI-generated wildlife video, including animals, dusty ground and camera movement. His assessment favors output quality while calling for lower resource consumption and latency.
The description includes a GPU rental discount link and coupon.
