On this page

Read the final response metrics

Ollama responses can include prompt token count, cached prompt tokens, output token count, model load time, prompt evaluation time, generation time, and total duration. Streaming responses provide those fields in the final chunk when generation is complete.

Those numbers let you separate a slow model load from a slow prompt evaluation or slow token generation. That is much more useful than one overall stopwatch.

Track context as the local limit

For a local model, the practical limit is often the context window and available RAM or VRAM. A long conversation can push context higher, reduce speed, and increase memory pressure even though there is no provider bill.

Watch loaded model size, context used, unload timing, and tokens per second. If performance falls, shorten the session, choose a smaller quantization, or unload models that are no longer needed.

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

Make the metrics visible

You can collect the API response fields in your own client. Super Notchy also detects local Ollama models and can show memory, context, load state, and optional generation measurements at the Mac's edge.

Treat local usage as capacity planning. The question is not “How many prompts remain?” It is “Can this model complete the next task at an acceptable speed without crowding the machine?”

Sources and further reading