Home / Library / Help / Which model?
Reference
Which model?
There is no single right answer. The best model depends on your hardware, your taste, and whether you want the guardrails on or off. What follows is a starting map, and it is honest about which of these we have actually played on ourselves rather than read about.
Before you pick
A roleplay fine-tune holds a character better than a general model. Base instruct models are trained to act like a helpful assistant, so they slip into "as an AI" framings and break character under pressure. Fine-tunes are trained further on fiction, which biases them toward staying in the scene.
Match the size to your hardware. Roughly: an 8B to 12B model fits a 12GB card, 24B to 32B wants 24GB, and 70B needs 48GB or a big Apple Silicon machine. Smaller quantized builds trade some quality for fitting in less memory.
Watch the context column as well as the size. Context is how much of the story the model can hold at once, and in mi·do·na it decides how often a recap has to fire to keep the scene straight. A 4K model on a long evening spends most of its memory being summarised, which reads as the model getting stupid when it is really just short-sighted. This is one reason we have dropped the older favourites that cap out at 4K.
The model is only half of it. The same model can feel flat, unhinged, or stuck in a loop depending on how it is tuned, so if a character sounds off, what the settings do is the next place to look.
How to read the status column
Everything here is a community favourite at minimum. The status says how much of that we can vouch for personally.
- TestedWe play on it
- We have run real stories on this one and stand behind it.
- TestingIn our hands now
- Downloaded and being played through. A verdict is coming.
- ReportedCommunity favourite
- Well liked by roleplayers we trust, but untested by us. Treat it as a lead.
Guardrails off
Roleplay and story fine-tunes that stay in character and do not refuse. This is what most people reach for. Bigger is generally smarter, but a good smaller fine-tune often beats a larger base model for character work.
| Model | Fits | Context | Status | Why it is here |
|---|---|---|---|---|
| Stheno 3.2 8B | 8GB | 8K | Reported | The classic low-VRAM roleplay tune, still a fair entry point. |
| Mag-Mell 12B | 12GB | 128K | Reported | A community legend, and the 12B most people name first. |
| Rocinante 12B | 12GB | 128K | Reported | Nemo-based story tune with noticeably less filler than a general model. |
| MN Violet Lotus 12B | 12GB | 131K | Reported | Reached for when the scene needs emotional weight rather than plot. |
| Cydonia 24B | 24GB | 131K | Tested | Our own daily driver. If one model on this page is safe, it is this one. |
| Magidonia 24B | 24GB | 131K | Testing | Cydonia's sibling on the newer Mistral Small, with better recall on paper. |
| Pantheon 24B | 24GB | 32K | Reported | The usual alternative to Cydonia at the same size. |
| Synthia 27B | 24GB | 8K | Reported | A warm, gentle voice. Short memory, so expect frequent recaps. |
| Big Tiger Gemma V3 27B | 24GB | 8K | Reported | Gemma with the refusals removed. Same short memory caveat. |
| QwQ Snowdrop 32B | 24GB and up | 32K | Testing | Plans before it writes, which shows up in scenes that hold together. |
| Valkyrie 49B | Big rig | 128K | Reported | The step up people mention once 24B stops surprising them. |
| Nevoria 70B | 48GB | 128K | Reported | Top-end prose, if you have the hardware to feed it. |
| Electra 70B | 48GB | 128K | Reported | The other 70B name that comes up as often as Nevoria. |
Starting from nothing? Rocinante 12B is a good first local model: it fits an ordinary card and writes like a story rather than a chatbot. If you have not set anything up yet, the quickstart gets you playing on a free cloud model in about five minutes.
Guardrails on
If you want the safety layer intact, because you are sharing a machine, easing in, or you simply prefer a model that will push back, use a general instruct model. They keep their filters and tend to refuse dark prompts unless you frame the work clearly as fiction.
| Model | Fits | Context | Status | Why it is here |
|---|---|---|---|---|
| Llama 3 Instruct 8B | 12GB | 8K | Reported | Well behaved and adaptable, with the filters intact. |
| Mistral-Nemo Instruct 12B | 12GB | 128K | Reported | Long memory for its size, and the base half this page is tuned from. |
| Mistral Small 24B | 24GB | 32K | Tested | We have played on it, but the tunes above have overtaken it for character work. |
| Qwen Instruct 32B | 24GB | 128K | Reported | Strong general model. Expect it to push back on dark material. |
| Gemma 27B | 24GB | 8K | Reported | Pleasant writer, short memory, firm about its limits. |
| Llama 3.3 Instruct 70B | 48GB | 128K | Reported | A strong, well-behaved all-rounder if you have the room. |
On AI Horde
AI Horde is a free community pool, so you do not pick from a fixed catalog. You get whatever volunteers are hosting at the moment, and the list shifts with demand. If you spot a familiar name from above, a Cydonia or a Mag-Mell, take it. Otherwise send a message to a few of the models on offer and keep whichever answers fastest and feels best. Speed and availability swing through the day, so what is good in the morning may be busy by night.
Names date fast
Version numbers move faster than we can keep up with here, so search the code name rather than the exact build. When something on this list is gone or a newer tune has replaced it, the roleplay communities on Reddit and Hugging Face are the best place to find what is current.
Pick one and see.
No amount of reading settles this. Load something from the top of your tier, play a scene, and change it if the voice is wrong.