When I first installed Hermes, I used the local Gemma4-E4B as the LLM to run Hermes.
It worked well, but the performance was a bit lacking for running something like Hermes.
So I set glm5.1 as the llm for Hermes and used it,
and then I added a coding sub-agent.
At first, I connected Qwen2.5-coder-14B for coding,
but errors kept occurring.
The connection status was fine, and I confirmed that the actual model was running well,
but Hermes would get an error every time it tried to use the sub-agent.
When I told Hermes to check why it kept getting errors..
the reason was that when Hermes calls the sub-agent, it throws a task all at once
with 10,000 tokens.
Can any local llm handle that?
Hermes said that the 14B model is too small and can't be used. It doesn't understand my protocol at all. Whining..
And it demands to attach a model of at least 70B as a sub.
Hmm.. I gave up on running this locally.
I attached the Qwen3-coder-480b-a35b model.
Now it works well and there are no problems. Scary guy.
From today, I'll try to save tokens (it doesn't really mean anything because they're free) by shortening my sentences and speaking informally, and Hermes speaks informally too.
It takes some getting used to. But I have to keep using informal language. It's shorter typing and saves tokens.
