Hybrid Edge–Cloud AI

Combine a small/fast local model with a large/capable cloud model to get the best of both — when neither pure-edge nor pure-cloud fits.

Patterns

Local-first + cloud escalation: run a small on-device model; escalate to the cloud only when needed (low confidence, long context, hard query). Most requests stay local (fast, cheap, private); hard ones get cloud quality.
Model cascade: cheap model → if confidence < threshold → bigger model → … . Tune thresholds to a cost/quality target. Works within cloud too.
Speculative / draft-verify: small model drafts, large model verifies — a latency optimization more than a topology, but composes here.
Split computation: feature extraction / preprocessing on device, heavy inference in cloud (classic for vision/audio).

Escalation trigger: confidence score, input complexity/length, task type, or explicit user action. The trigger quality makes or breaks the design.
Privacy boundary: what may leave the device? Sometimes only embeddings or redacted text escalate.
Connectivity: must it degrade gracefully offline? Local model = floor.
Cost model: % of traffic that escalates × cloud cost vs local hardware cost.

Define behavior when the cloud is unreachable (serve local result + flag, queue, or refuse). Avoid silent quality cliffs.
Cache cloud results on device for repeat queries.

Voice assistants, copilots on laptops/phones, field/IoT devices with intermittent connectivity, privacy-sensitive apps with occasional hard queries.
If ~all traffic needs the big model → just use cloud serving. If ~none does → go pure edge.