Generation stats
Every finished assistant reply shows a small metrics chip under the bubble, next to its … menu. It gives you a quick read on how the model performed without leaving the conversation.
What you see
The chip reads like ⏱ 42.3 tok/s · 512 tok · 12.10s:
- Tokens per second - how fast the reply was generated. Useful for comparing models on your hardware at a glance.
- Tokens - the completion tokens of this reply; the prompt isn’t counted. When the model used tools, every round of the reply is added up.
- Elapsed time - wall-clock duration of the whole reply, from the moment the request goes out to the last token, including model loading and any tool calls.
The chip appears as soon as the reply finishes and is saved with the message, so it stays put after a restart and travels in data backups.
Where the numbers come from
- Ollama - reports the token count itself, and the rate uses Ollama’s own generation time, so model loading and prompt processing don’t drag it down.
- OpenAI-compatible - the count comes from the endpoint’s usage report when it sends one; it follows the reply’s last chunk, and Loach reads on for it (up to 2 seconds on a server that keeps the stream open). Older proxies ignore Loach’s request for it; Loach then counts the streamed chunks as an approximation. The rate is worked out from the wall clock.
- Approval waits - time a reply spends paused on a tool approval card is left out of the rate but still counted in the elapsed time.
If the numbers don’t match your provider’s dashboard, these differences are usually why - see Troubleshooting.
Stopped and failed replies
The numbers arrive only when a reply completes. A reply you stop, one that Respond now pre-empts, and one that ends in an error - including one the server cuts off mid-reply - keep their partial text but get no chip. See Concurrency & queue for details on how partial output is preserved.
Totals for a chat
Type /stats in the composer for a summary of the open chat: the message count (user and assistant), the assistant tokens across every reply that has a chip, the average tokens per second, the last reply’s numbers, and the model and provider. The average divides the total tokens by the total elapsed time, so on Ollama it reads lower than the per-reply rates.
Private Chat shows a shorter line - tokens and tokens per second - under each finished reply, and like everything else there it is never stored.