---
title: "Model advice - Feedback - Yumina Community"
description: "So I've been playing for a while. Still new, but I've been wondering if I'm making good use of the available models. So far, I've tried the free model and a couple others, Claude 3 Haiku being my…"
url: https://yumina.io/app/community/thread/c61e4726-0a08-4c64-aca8-bb1d51b8b475
lang: en
---

[Community](https://yumina.io/app/community) · [Feedback](https://yumina.io/app/community/feedback)

# Model advice

[Lasinn Vargr](https://yumina.io/@lasinnvargr) · 2026-09-06

So I've been playing for a while. Still new, but I've been wondering if I'm making good use of the available models. So far, I've tried the free model and a couple others, Claude 3 Haiku being my default, normally. Since I do play for long periods, I want to stay in the budget zone. However, I've noticed that AI has a tendency of forgetting characters, impose emotions on my character, or even take control and tell me what I do without my input. Pausing the rp for a moment to remind it about characters and instructions seems to fix it for a while, but after enough replies it just needs a reminder again.

I've also noticed that some models, especially the free model, have this... way of narrating that repeats itself a lot. "That's not X, it's Y". "this thing is X. Y. Z." "This character moves like someone who \[insert background reference\]" As well as really liking some specific words and using them everywhere. It'd be ok if it wasn't so... common, it gets repetitive.

Is this common with every AI model? Which ones work better for you? And also, if I switch models mid session, like using the free model to give instructions or corrections, and then switch to Haiku to keep playing, would it run the risk of corrupting anything or the sort?

## 5 replies

1. [Yoshishi](https://yumina.io/@yoshishi) · 2026-09-06 I personally use Deepseek 3.2 (128k context size) as a fellow budget person... (It's better than Claude 3 for my experiences, and follow prompts really well. Even thinks~), the free model is mid... No, switching models mid session (unless context size exceeds the models capacity) doesn't corrupt anything. (Also, use things like an (OOC:) block chain at the very bottom (like copy and paste it, with having instructions you want it to do... mine is usually 2k tokens long, but that's just me.
2. [Lasinn Vargr](https://yumina.io/@lasinnvargr) · 2026-09-06 2k tokens long and you're using budget? Gawd damn...
3. [🐱](https://yumina.io/@stay0325) · 2026-09-07 Have u activate the memory extension? And if the card that u r currently playing has a presets take up a huge chunk of tokens, it's better u don't use free model or else u gonna regret it for sure
4. [Hannah Berryblue](https://yumina.io/@hannahberryblue) · 2026-09-07 Es que la verdad el Claude Haiku sí está bien feillo ¿Para qué decirte que no? Intenta el DeepSeek. Editado por las dudas: Y no te olvides de todos modos que, hay juegos en los que no obtendrás la experiencia completa a menos que los juegues con modelos Gemini azul.
5. [大大爆肝](https://yumina.io/@zxyw) · 2026-09-07 Gemini 3应该是综合性价比比较好的模型，听话，能处理复杂指令，就是有的时候会吃设定，脱离卡本身框架（主要你指令别太有强制性）。
