Provider Selection Guidance

Narrow model choice by picking a provider first, then capability tier, then model size within that provider's lineup — hosted API or self-hosted.

Once model selection criteria narrow what's needed, the usual sequence is: pick a provider based on requirements, features, and a scrappy-versus-premium preference; narrow further by capability; then choose model size within that provider's lineup. Providers differ in emphasis — some lean toward alignment and safety research, some toward highly efficient open-weight models suited to specialized configurations, some toward strong retrieval-augmented-generation performance, some toward ecosystem integration and large-scale infrastructure, some toward large, highly capable open-access models — and the field shifts fast enough that any specific ranking goes stale quickly. Model-comparison sites are a reasonable starting point for prototyping and narrowing the field without committing early.

Open-source options avoid depending on a hosted LLM-as-a-service provider entirely, at the cost of real hosting effort — platforms that manage inference infrastructure for you ease this, but self-hosting is only worth it once an application is large enough to justify the infrastructure investment. An agile middle path: prototype on an easily accessible hosted API, and only move to a different platform (self-hosted or otherwise) before going to production if the economics or requirements demand it by then.

Within a chosen provider, model-size choice trades completion quality against cost and possibly latency; the default heuristic is to pick the smallest model that reliably delivers on the task. One counterintuitive tip: prototype with a slightly larger, pricier model than you think you can afford. Flagship releases tend to push older models' prices down over time, so by the time an application reaches public beta a better model may well be affordable — and having prompt engineering and postprocessing already tuned against that stronger model pays off immediately once it becomes affordable, rather than needing a re-tuning pass later.

Sources

Prompt Engineering for LLMs (John Berryman), ch. 7