Use local AI when the task is small and privacy-sensitive
Browser or device AI is a strong fit for classification, extraction, redaction, embeddings, and some summarization or transcription tasks. It can reduce round-trip latency, keep repeated use inexpensive, and avoid sending the input to an application server.
The tradeoff is hardware diversity. A model that feels instant on a recent laptop may be slow or unavailable on a phone, an older browser, or a managed work device. Downloads and memory limits must be part of the product design.
Use cloud AI when capability and consistency matter more
Cloud runtimes are appropriate when the model is too large for typical devices, when work must continue after the browser closes, or when multiple people need to share the same state. They also simplify support because the application controls the hardware environment.
That control introduces backend operations: protected provider keys, authentication, rate limits, logs, spend limits, job state, and data-retention decisions.
A hybrid route is often the product
Some applications can classify or redact locally, then send only approved material to a larger cloud model. Others can use a local default with an explicit “use a larger model” action. The route should be visible instead of silently changing behind the interface.
Start with the smallest honest route
If a 70 MB browser model solves the task, a GPU server may add complexity without adding value. If the task needs a multi-billion-parameter model and shared history, hiding that fact behind a “local” label is equally unhelpful.