County 311 needs high-volume address-normalization suggestions with tight cost and latency budgets. Which generative model choice fits best in Foundry?

A small language model sized for high-volume, low-latency address normalization best fits County 311's cost and latency budgets in Foundry. Address tidy-up is a repetitive, structured generative (or generative-assisted) task: standardize street suffixes, fix casing, and suggest canonical forms at scale. Smaller models typically offer lower per-call cost and faster responses, which matter when every ticket or chat turn may normalize an address under tight SLAs.

A large reasoning model tuned for high-volume, low-latency address normalization still burns budget and latency on reasoning depth that address normalization does not require, so it cannot match a small model's cost and speed at 311 volume. A small language model with multi-step reasoning for high-volume address normalization adds unnecessary reasoning steps for repetitive short suggestions, increasing per-call latency and cost versus a plain small model sized directly for tidy-up. A multi-agent workflow tuned for high-volume, low-latency address normalization still adds coordination overhead between agents for a task a single small model can already handle directly within the latency and cost budget.

Exam caveat: model selection questions trade capability against cost/latency; prefer the smallest model class that meets quality for the municipal workload. Operational check: benchmark p95 latency and unit cost for address-normalization prompts on a small Foundry model versus a large reasoning model SKU, and promote the small model if quality gates still pass on a labeled 311 address sample.