Best models and tips to make a local LLM sound human?
Posted by LongjumpingHeat8486@reddit | LocalLLaMA | View on Reddit | 6 comments
Hey everyone,
I’m running a local instance (right now I'm thinking llama3.2 or dolphin-llama3) and I want it to interact with users naturally. Right now, it either sounds too AI-like (obviously).
I have a few questions, which local models are the best for natural casual conversation whilst listening to guidelines? Since I notice most models will go completely out of their restrictions and start spewing paragraphs of random stuff. Are there any good tricks to make the LLM sound more human, like slang, casual phrasing or context awareness? And how do you handle proactive messages without flooding or sounding robotic? Any tips prompts or model recommendations would be MASSIVELY appreciated.
Thanks so much in advance!
draconisx4@reddit
To get your local LLM chatting more naturally while sticking to guidelines, focus on prompt engineering with built-in safety checks to prevent it from ignoring restrictions. Always prioritize runtime oversight so it doesn't spiral out of control during interactions. Start simple: test with varied prompts and monitor outputs closely to balance humanity with safety.
LongjumpingHeat8486@reddit (OP)
This is super helpful thank you!
FridaGerman@reddit
um..what was super helpful? maybe PM please
LongjumpingHeat8486@reddit (OP)
They were just talking about models built for role play type stuff so they're finely tuned to reply somewhat like humans! :]
WhoRoger@reddit
You didn't specify size, but under 10B I enjoy LFM2 (1/8B), Smollm3 (3B) and Granite 4 (1/7B).
They all have pretty normal default personalities, and respond well to normal instructions in the system prompt (act casual, be funny...). It's a simple transition if you're used to Llama models.
Also look for Hermes merges. Hermes are older datasets with more human-like personality, and people often plug them into other models to inject more human default behaviour. So there are all kinds of hybrids like DeepSeek-Hermes, Qwen3-Hermes, Granite-Hermes, and they come in lots of sizes.
Base Qwen and Gemma are even more malleable with system prompts, but I find them harder to manage since it's way too easy for them to go totally off the rails.
Witty_Mycologist_995@reddit
Use something like Gemma 3/4 or GLM. Also any LLM that’s tuned for roleplay will sound more human.