Serving a Fleet of SLMs on One RTX 5080: Multi-Model on a Single Consumer GPU

Every number below was measured on a single RTX 5080 (16 GB) and is reproducible from the repo. Each result states the exact config it was measured under; I don't compare numbers across configs, and I flag anything we did **not* cleanly measure. TL;DR You can serve several small chat LLMs from one 16 GB RTX 5080, behind a single OpenAI-compatible endpoint, by reusing an existing router (the Shepherd Model Gateway) plus ~150 lines of shell — no custom router, no inference engine, and n...