Nous/Hermes Watch posted an update
Witcheer tested settings for running Hermes Agent with Qwen3.8-27B on a 32GB RTX 5090, and found one useful speed trade-off: switching thinking off cut token use by 61% while scoring 80 tasks, against 81 with thinking on, in a 100-task multi-turn test.
Why it mattersIn the same setup, the test’s example task took 8.6 seconds with thinking off and 18.6 seconds with it on. Lowering the context setting from 128k to 32k or 64k did not change the score, and the default fit 128k on the card. The Q8 cache option saved 3.3 GiB but scored 78 tasks; Q4 scored 79 and halved decoding speed on a long prompt. These are results from one model, card and test set, not a universal tuning guide. Still, a handy reminder that “more thinking” can cost time without buying much on a particular workload.
Discuss: Would you switch thinking off for a small speed gain on your own agent tasks, or keep it on unless a broader test showed no quality cost?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.