Bundling a quantized model into an app is a weekend. Keeping 200,000 installs on a compatible (app, runtime, weights) triple for a year is the actual work. The model is not a library you link once; it is a 1.5–2 GB artifact that ships on a schedule you don't control, to hardware you've never held, over networks that drop mid-transfer.
I've watched more edge-AI incidents come from artifact lifecycle than from model quality.
Top comments (0)