Generation stats
Every assistant reply shows a small metrics chip under the bubble. It gives you a quick read on how the model performed without leaving the conversation.
What you see
The chip reports, when the provider makes the data available:
- Tokens per second — the live throughput of the completion. Useful for comparing models on your hardware at a glance.
- Total tokens — prompt plus completion. Helps you reason about context budgets.
- Elapsed time — wall-clock duration of the request, end-to-end.
When some numbers are missing
Not every provider returns every metric. Local Ollama reports all of them. Hosted OpenAI-compatible endpoints typically return token counts but not always tokens per second; Loach renders whatever was given and silently omits the rest. The elapsed time is always shown because it is measured locally.
Cancelled and errored replies
If a generation is cancelled or errors mid-stream, the chip still shows the partial token count and elapsed time so you can tell how far the model got before it stopped. See Concurrency & queue for details on how partial output is preserved.