Suggested Text Models
Sometimes people assume due to the scale of Yozakura, generating memories and managing relationships between dozens of characters, that an especially strong model is required. Fortunately, that's not the case, since characters effectively only interact with a few other characters at a time (usually just one other character, depending on your group chat settings). Users have reported Yozakura working well with models as small as 4B.
Smaller models are more likely to suffer from hallucination, misattribution, forgetfulness, etc, but just allowing this to happen often provides a better experience than one might expect. Alteratively, since character memories are easy to edit, a user who is open to doing some work themselves can correct and improve character memories by hand (this also enables steering characters in the user's desired direction, which can be entertaining even when you have a strong model).
Below are some local models I've tried or have heard reports about (and if you find a particularly good model, please share your report too, you can contact via the links page).
| Model | Size | Notes |
|---|---|---|
| Gemma 4 26B A4B | 26B (but see notes) | This is my daily driver now and I find it radically smarter than 12B models. Due to the MoE achitecture, this model has pretty good performance with as little as 8GB VRAM (though it will likely feel slow at the memory generation phase). My favorite finetune for Yozakura so far is hiwaifu-research/WaifuGemma4-26b-a4b-v1. Since I have done most of my testing with Gemma 4 26B A4B, the default settings in Yozakura are well tuned for it. |
| Gemma 4 12B IT | 12B | It's not as strong as the 26B model, which is to be expected, but it might be the best 12B model for coherence. At the same time, it might be a little less creative and entertaining than some other 12B models. |
| NemoMix Unleashed 12B | 12B | This was my clear favorite 12B model before Gemma 4, providing a good mixture of believable character behavior and memory correctness. It can be funnier than Gemma 4. But its biggest advantage over Gemma 4 (including the 26B model) is that it's a lot less reflexively agreeable. This makes conflicts between characters more believable. However, compared to Gemma 4 26B A4B, it's much dumber at generating memories. |
| Mag Mell R1 | 12B | Another 12B model that's a long-time favorite in the RP community. It can be pretty aggressive and I find it a bit much, but it can be hilarious sometimes in a way that other models rarely match. Worth a shot if you value spontaneity, but hard to recommend broadly. |
| Llama 3.1 8B Instruct | 8B | A workhorse model. A little plain, but does a good job mechanically. There are a lot of finetunes out there that might give a more interesting experience, but some of the RP-specialized ones do poorly on generating memories and sometimes fail to understand that they're supposed to be generating memories and instead try to continue the conversation or something. |
| Gemma 4 E2B | 5B | Yes you can use this tiny model for Yozakura and it's surprisingly good. From what I saw in my limited time with it, I think it's at least as good as Llama 3.1 8B, and may even approach 12B models. You may not even need a GPU for this model, many machines can get acceptable performance purely from CPU inference. |
I haven't used big cloud models much, but Deepseek V4 Flash seems to be the most popular one at the time of this writing.