Performance

The performance target is predictable low memory rather than maximum request throughput. The gateway bounds request bodies and concurrent work and streams responses without retaining the full generation.

Always benchmark the exact provider, concurrency, request size, and container limit used in deployment.

The benchmark harness also records CPU, memory, network I/O, block I/O, process count, swap, and OOM/restart state for Go Feather Route and the pinned LiteLLM comparison image.