These days I'm making a simple game with Hermes.
Instead of just playing games, I'm playing by making games.
I'm building it using the Godot engine by connecting a commercial LLM to Hermes.
Since I've been using LLMs cheaply, models without multimodal, i.e., vision capabilities, have been my main focus.
So I connected Gemma4-26B-A4B, which has multimodal capabilities, to Hermes
and tried it out a bit. I was quite disappointed.
The model that seemed quite smart when given small tasks, but as the number of file types increased and the scale grew,
it immediately became confused and only created more errors in the tasks I asked it to solve.
I didn't know about the limitations of 30B-class models until I gave it large tasks.
After that, my enthusiasm for local LLMs cooled considerably.
Of course, having used 1T-class models and then trying a 26B model, it's natural that I'm not satisfied,
but it seems weight class differences are inevitable.
