Unswarm: a swarm of models, one GPU at a time

Field notes

Unswarm routes model requests through a single OpenAI-compatible endpoint and loads only the requested container, allowing several models to share a small GPU rather than answering concurrently. Its stated feature set extends beyond swapping: conversation affinity can prevent mid-chat reloads, while router profiles provide fallback ordering. Separate machines connect through outbound WebSockets, and the page lists live GPU, CPU, RAM, latency, token, and cost telemetry.