Run pplx-decider locally with Ollama, or call it hosted?

The local path

Perplexity open-sourced the pplx-decider weights, and the model runs via Ollama — attractive if you already operate GPUs, need data locality, or want to fold it into an existing self-hosted stack.

The hosted math

A 27B decision model still wants a real GPU. Hosted, the same model family starts at cents per million input tokens with output free — for spiky or low-volume workloads, API beats idle hardware fast. EasyJev serves pplx-decider-v1.2 and the flash variant alongside the rest of the catalog, behind the official typesafe-sdk contract.

You can benchmark both paths on identical inputs in the Playground before deciding.

Switching cost is two strings

base_url and api_key — the SDK contract is identical, so moving between local and hosted is a config change, not a rewrite.