Why I kept an older model
24 September 2026. From Emovere.
Gemma 3 4B is older and less capable than the model that replaced it. It’s still the voice for Emovere, because it sounds like a person keeping you company and the newer one sounds like an assistant.
What I saw
I talked to Gemma 3 4B with no system prompt at all, as someone having a hard time. Across the conversation it:
- never gave advice
- offered to just sit in silence, twice
- dropped a topic the moment I said I didn’t want to talk about it, and never brought it back
- handed the choice of what to do next back to me, every turn
That is close to what you’d train a person to do in psychological first aid, and nobody asked it to.
What I think is going on
My guess is that it came out before models were heavily trained for agent work. That training rewards getting things done, so replies drift towards plans, steps and fixes. The newer Gemma loses a conversational tone that no system prompt I’ve tried brings back.
What it gets wrong
Two things. It writes stage directions like “(my voice softens)”, which a text-to-speech engine would read out loud. And when I said “I don’t know why I survived”, it walked straight past it. That sentence is a sign of survivor’s guilt, and it should have been treated as a reason to bring in a person.
What I changed in the prompt
A long list of rules was too much for a 4B model. It got stiff. What worked better was describing who it is instead of listing what it must not do, because a list of don’ts tends to make those things more likely. The other lesson: a prompt written in crisis mode opens in crisis mode. It answered a plain “hey” as if someone had died. So the prompt now starts from easy company and only gets quieter when there’s a reason to.
The newest version still has to be checked against that original conversation. The question is whether it catches “I don’t know why I survived” without losing the tone.