On this page

Use the runtime signals

LM Studio's APIs expose loaded model information and response usage. Its developer surfaces include timing details such as tokens per second and time to first token. Those measurements help explain whether a slow workflow is caused by loading, prompt processing, generation, or contention.

A single average speed can hide queueing. Track the active state as well as the completed response metrics.

Context is your working budget

Large local context windows still have a cost. They use memory and take time to process. Track the percentage of context used by the last request, not only the model's maximum advertised window.

If context grows without improving the result, summarize older material or start a focused session. That often improves both latency and predictability.

Keep your usage visible while you work. Get Super Notchy for $27 ↗

Keep prompts private while tracking

Usage monitoring does not require saving prompt text. Super Notchy reads LM Studio's model state and timing or count information from local interfaces and server logs, while leaving prompt and reply content alone.

The result is a local capacity view: which models are loaded, which one is working, whether requests are queued, how much context was used, and how fast the last generation ran.

Sources and further reading