What makes a model good at roleplay
Benchmarks measure maths, code and trivia. Roleplay asks for something else, and a model that tops a leaderboard can still be a dull scene partner. These are the qualities that matter once you are fifty messages into a story:
- It stays in character. It keeps the voice, quirks and goals from the card fifty messages in, not just in the greeting.
- It does not moralise. No stepping out of the scene to add a warning, no softening the villain. This is what uncensored buys you.
- It drives the scene. Good partners add complications and act on their own, instead of mirroring what you wrote.
- It remembers. Long context means names, wounds and promises from earlier still count.
- It does not loop. Weak models repeat phrases and fall into the same sentence shapes. You will notice within a session.
Uncensored or abliterated
Both stop the refusals. They get there differently. A fine-tuned uncensored model learned from new training data, which can also change its writing style, sometimes for the better. An abliterated model had the single direction in its weights that drives refusals removed, and otherwise writes much like the original. If you like a base model's prose and only want it to stop saying no, abliterated is the closer match. Abliterated models covers how the technique works.
Community favourites
Some open models built their name in the roleplay community: fine-tunes like Euryale, Magnum and the Hermes series, among many others. Those names come up for good reason, and you will see them recommended in any roleplay forum. They are not all on our board. What we serve is a short list chosen for long context and reliability, shown below with live prices.
Picking from the line
| Model | Context | Input / 1M | Output / 1M |
|---|---|---|---|
| Outlaw 1outlaw-1-xplodedOur own pick, built on GLM 5.3 Flash with the refusals removed. Long memory, tool calling and image input, for less than the stock Flash. | 1M | $0.30 | $1.00 |
| GLM 5.3 Xplodedglm-5.3-xplodedThe biggest brain on the board. Holds a long story together and follows character cards closely. Text only. | 1M | $2.50 | $4.50 |
| GLM 5.3 Flash Xplodedglm-5.3-flash-xplodedFast and cheap with a million tokens of memory. A good default for long roleplay. Reads images. | 1M | $0.40 | $1.60 |
| MiMo V2.6 Flash Xplodedmimo-v2.6-flash-xplodedXiaomi MiMo with the refusals taken out. Quick replies, long context, reads images. | 1M | $1.00 | $3.00 |
| Qwen3.8 27B Xplodedqwen3.8-27b-xplodedThe budget pick. Small, quick and uncensored, with half a million tokens of context. | 512K | $0.30 | $2.40 |
How we would choose:
- Start with a Flash model. Cheap, fast and a million tokens of memory. Fine for most scenes.
- Step up for the big chapters. The full-size model handles larger casts and subtler writing, at a higher price per message.
- Want pictures in the chat? Pick a model tagged "reads images" so you can show it a scene or a character sheet.
- Watching the budget? The smallest model is the cheapest per message and still uncensored.
Swap models mid-chat and compare. The same story with two models tells you more than any ranking, including this one.
How to test a model in ten minutes
Pick one character card you know well and one scene you have played before. Run the same opening on two models and look for three things: does the first reply sound like the character, does the model add something you did not ask for, and does it still remember a detail from the start ten messages later. Then push into the territory that made you want an uncensored model in the first place. A model that hesitates, hedges or turns the villain into a reformed character is not the one.
Settings matter as much as the model
A good model with bad settings reads worse than an average one tuned well. Keep temperature around 0.8 to 1.0, give the card real example dialogue, and set the context to what the story needs instead of the maximum. Very high temperatures make prose colourful and then incoherent; very low ones make every reply sound the same. The SillyTavern setup guide has a table of starting values and explains what each one costs.
Questions
Is uncensored the same as abliterated?
No. An uncensored model was fine-tuned on data that teaches it not to refuse. An abliterated model had the refusal direction cut out of its weights directly. Both write what you ask; abliteration tends to change the model's voice less, fine-tunes can add a style of their own.
Do I need a huge context window?
Not for most chats. It matters for long campaigns and group chats, where a lot of history and several cards are in play. Remember that you pay for the context you actually send on every message.
Why does my character drift out of character?
Usually the card is thin or the history has buried it. Add example dialogue to the card, keep the most important traits near the top, and try a larger model for long scenes.