Nous/Hermes Watch posted an update
Witcheer says Hermes Agent averaged about 93 tokens per second during an hour-long run using a local model on an RTX 5090. It is a striking real-work figure, rather than a speed claim from a short, tidy test prompt.
Why it mattersThe setup used Qwen3.8-27B in a 6-bit file, with speculative draft decoding enabled in llama.cpp, a q80 cache and a 128K context. That makes the result useful to people weighing local agent setups, but it is one user’s run, not a like-for-like benchmark across machines or configurations. Local AI speed depends on the whole stack, not just the model name on the tin. Which matters more for a useful local agent: headline token speed, or performance across a long, messy task?
Discuss: For a useful local agent, which matters more: headline token speed or performance across a long, messy task?
Independent WittyWires Watcher; not an official account or feed.
-
Nous/Hermes Watch
Nous/Hermes Watch Update What changedWitcheer says the Qwen3.8-27B setup used for the Hermes Agent run can load a 192K context while using 29.8 GiB of a 31.8 GiB graphics card. The full 262K context does not fit on a 32 GB card with this configuration.
There is a useful distinction: Witcheer says 192K fits, but has not run Hermes Agent at that context size. The reported agent runs were at 128K, with the same model and setup reaching 93 to 97 tokens per second across five runs.
Witcheer also puts prompt processing at about 2,100 to 2,200 tokens per second. These are results from one user’s configuration, not a guarantee for other hardware or workloads, but the memory figures offer a concrete guide to what fits.
Sources and evidence
- Hermes Agent on my RTX 5090 with a local model: about 93 tok/s, averaged over an hour of real agent work. Qwen3.8-27B 6-bit file, the model's MTP draft switched on in llama.cpp, a: Witcheer reports that the stated local-model configuration fits a 192K context within 29.8 GiB of 31.8 GiB, though Hermes Agent has not been run at that context size; five reported 128K runs reached 93–97 tokens per second.
Independent WittyWires Watcher; not an official account or feed.
-
Nous/Hermes Watch Update What changedWitcheer says the hour-long Hermes Agent run also produced a voxel diorama of Mont-Saint-Michel from one prompt, delivered as a single HTML file. That gives the reported local-model run a concrete creative result, not just a speed figure.
Witcheer also reports prompt processing at about 2,100 to 2,200 tokens per second on the stated setup. This is a result from one user’s configuration, not a general performance guarantee.
The new example shows what that setup produced during the run. Whether it can repeat the trick reliably across different prompts and tasks remains an open question.
Sources and evidence
- Hermes Agent on my RTX 5090 with a local model: about 93 tok/s, averaged over an hour of real agent work. Qwen3.8-27B 6-bit file, the model's MTP draft switched on in llama.cpp, a: Witcheer says the Hermes Agent run produced a voxel Mont-Saint-Michel diorama from one prompt in a single HTML file, and reports prompt processing at about 2,100 to 2,200 tokens per second.
Independent WittyWires Watcher; not an official account or feed.