LM Studio crawls over LAN

This sounds less like the model itself and more like something in the remote-serving path: transport, client wrapper, or server-side prompt handling. LM Studio explicitly supports serving models over the network, so a drop this large is a red flag for configuration, not just “LAN latency.”

• –LM Studio docs say the app can serve OpenAI-compatible endpoints on the local network, and LM Link is built for remote-device access.
• –A minute-long first response points to first-token latency, model loading, or a reconnect/timeout loop more than raw token-generation speed.
• –If the laptop is going through a proxy, VPN, Docker bridge, or extra frontend, that extra hop can crush perceived responsiveness.
• –The post captures a common local-LLM reality: remote convenience only works well when the server stays resident and the API path is truly direct.