Outerport (YC S24) – Instant model weight hot-swapping

Outerport, a YC S24 startup, has built tooling for hot-swapping AI model weights during inference without requiring service restarts. The system enables dynamic model switching at runtime, allowing operators to change which weights are loaded in memory on demand.

For teams running inference infrastructure, this reduces operational friction around A/B testing, model rollouts, and resource allocation. Instead of deploying separate inference endpoints per model variant, operators can maintain a single instance and switch weights programmatically. This cuts infrastructure costs for multi-model serving scenarios and eliminates downtime during model transitions—a meaningful constraint for production systems.

The operational shift is concrete: model switching moves from a deployment operation (requiring restart cycles) to a runtime configuration change. Teams managing resource-constrained inference or running frequent experimentation workflows can reduce idle capacity and deployment latency. This particularly benefits builders using smaller instances where spinning up parallel endpoints isn't economical, and those optimizing for rapid A/B test iteration where switching speed directly impacts experiment throughput.