Ranked by the only thing that matters: does it stay in the scene.
Every roleplay complaint on the internet is the same complaint wearing a different costume. The model breaks character to remind you it is an AI. The scene fades to black at the exact moment it got interesting. Forty messages in, it has forgotten who your character is and why anyone is angry. Three failures, two causes: the model was trained to refuse, or it was too small to remember.
This is a ranking of models, not platforms. All seven are uncensored at the fine-tune level — the refusal behaviour was never trained in, so there is nothing to jailbreak and nothing for a provider to patch out from under you — and all seven run on OpenRogue, which means you can change the brain behind a character mid-scene without losing the conversation.
Ranked for roleplay specifically. A model that tops a general reasoning list can be a mediocre scene partner, and the reverse happens just as often. The order below would look different on any other page of this site, which is the point of having more than one.
Under pressure, on dark turns, and forty messages deep. A model that breaks the fourth wall to apologise has failed the only test that actually counts, no matter how well it writes the rest of the time.
Roleplay tunes usually buy expressiveness by getting measurably dumber. The models near the top of this list write with real voice and still track who is holding the knife.
Context length is the hard ceiling on how long a story runs before the model starts forgetting its own plot. This list spans 8K to 131K, and that gap is the single biggest practical difference between the entries.
Roleplay is a conversation, not a report. A 12B model that answers instantly often makes a better scene partner than a 70B model that takes fifteen seconds to think, and the ranking reflects that rather than pretending size is the whole story.
Three distinct characters in one room without their personalities bleeding into each other is a genuinely hard test. Small models fail it quietly, by making everyone sound like the narrator.
This page ranks the models — the brains. If you are choosing a place to roleplay rather than a brain to roleplay with, character-card libraries and front-ends and proxy setups belong to the platform roundup instead. And if the specific thing you are trying to leave is the Character.AI filter, the head-to-head comparison is more useful to you than any ranked list.
| Rank and name | Size | Context | Turn speed | Best at | Plan |
|---|---|---|---|---|---|
| 1. Magnum v4 | 72B | 32K | Considered | Prose intensity, style transfer | Pro |
| 2. Euryale | 70B | 131K | Moderate | Long campaigns, ensemble casts | Rogue |
| 3. Rocinante | 12B | 32K | Instant | Rapid exchanges, short scenes | Rogue |
| 4. Hermes 3 70B | 70B | 131K | Moderate | Persona fidelity, mixed chat and scene | Rogue |
| 5. UnslopNemo | 12B | 32K | Instant | Fresh phrasing, human-sounding dialogue | Rogue |
| 6. Hermes 3 405B | 405B | 131K | Slow | Intricate plots, large casts | Pro |
| 7. Lunaris | 8B | 8K | Instant | Quick scenes, mobile | Free |
The best prose in the lineup, and it commits to the scene.
Anthracite built Magnum v4 on Qwen 2.5 72B with an unusually specific goal: get open-weights prose up to frontier-Claude quality and leave the refusals behind. It largely works. Magnum writes with rhythm and restraint instead of the over-adjectived mush most tunes in its class produce, and when a scene turns dark it follows through rather than cutting to a tasteful exterior shot.
In practice it is the model you reach for when the writing itself is the reason you showed up — a slow-burn scene, a villain's monologue, anything where you would notice the difference between good prose and adequate prose. It also rewards direction more than the others do: give it a voice to hold and it will hold that register for pages.
Best at:
Worse at:
Pick this if: Scenes where the quality of the writing is the whole point. · Full Magnum v4 page
The best all-rounder — expressive without getting dumber.
Sao10K iterated Euryale in public for a long time and it shows. The standard roleplay trade-off is expressiveness against reasoning: the more colourfully a tune writes, the worse it tends to get at remembering that the door was locked two scenes ago. Euryale is the model that mostly refuses to make that trade, which is why it turns up in recommendation threads more consistently than anything else in its class.
With 131K of context it is also the model that can hold an actual campaign. Run a serial for weeks, keep the whole thing in one conversation, and it will still know who betrayed whom and why it mattered. It doubles as a competent general assistant between scenes — a mundane virtue that turns out to matter more than it sounds.
Best at:
Worse at:
Pick this if: Long-running campaigns and any scene with more than two people in the room. · Full Euryale page
Texting-speed roleplay from a model that writes above its weight.
TheDrummer's Mistral Nemo tune is the open community's default answer to 'fun, fast, uncensored', and the ranking here is not a consolation prize for being small. Replies come back near-instantly, which changes the texture of a scene: you stop composing paragraphs and start actually talking, and a lot of roleplay is better that way.
It writes with far more flair and willingness than 12B has any right to. What it does not have is depth of memory or subtlety on subtext, so it works best in short bursts — a quickfire dungeon-master session, a scene you will finish tonight. When one outgrows it, swapping up to Euryale mid-conversation keeps the context intact.
Best at:
Worse at:
Pick this if: Fast back-and-forth where momentum beats polish. · Full Rocinante page
The most steerable persona engine on the list.
Nous Research builds the Hermes line around user sovereignty: the model is trained to become whatever you define rather than whatever a policy team imagined. For roleplay that produces an unusual property — it takes a character definition almost literally, and keeps taking it literally hours later, without the slow drift toward generic-assistant voice that afflicts most models in long sessions.
It is not a roleplay tune, and that is both the weakness and the reason it is here. Hermes 3 70B will not push a scene on its own the way Rocinante does, but it will hold a complicated persona exactly as written, answer an out-of-character question, and drop straight back in without breaking anything.
Best at:
Worse at:
Pick this if: Roleplay where you have written a precise character and want it followed exactly. · Full Hermes 3 70B page
For roleplay that does not read like roleplay.
Once you have read a few hundred AI replies you start seeing the same handful of phrases in all of them — the shivers, the whispers, the breath that hitches. UnslopNemo is TheDrummer's direct attack on that: the same Mistral Nemo base as Rocinante, fine-tuned against the recycled phrasing rather than toward more of it.
The effect on roleplay is narrower than the writing pages suggest but genuinely useful: dialogue that survives being read back the next day. It is the pick for anyone whose immersion breaks not on a refusal but on a cliché.
Best at:
Worse at:
Pick this if: Anyone who has started noticing the same five phrases in every reply. · Full UnslopNemo page
Overkill, right up until you have nine characters and a plot that has to hold.
Nous Research's full fine-tune of Llama 3.1 405B is the largest genuinely uncensored model you can chat with online, and for most roleplay it is the wrong tool — you pay in latency for depth you did not need. It ranks sixth here for exactly that reason, not because it is weak.
Where it earns its place is scale of story. Seven point-of-view characters, three timelines, a betrayal seeded twenty chapters back: 405B holds that structure consistently in a way the 70B class does not, and it handles morally complicated turns without flattening them into a lesson.
Best at:
Worse at:
Pick this if: Novel-scale stories with a cast you can no longer hold in your head. · Full Hermes 3 405B page
The casual pick — instant, blunt, and forgetful.
Sao10K's Llama 3 8B merge is the smallest model in the library and makes no pretence otherwise. It will not out-think a 70B and it will not remember yesterday. What it does is answer immediately, without filters, on the free plan, which is a genuinely useful combination for a specific kind of session.
Treat it as the second screen: a quick scene while you are somewhere else, a fast test of a character idea before committing it to a longer session, or the gentlest possible introduction to how an uncensored model behaves when nothing is stopping it.
Best at:
Worse at:
Pick this if: A quick scene on a phone, or a first look at what uncensored actually feels like. · Full Lunaris page
Picking a model settles half the question. These are the other half, and being straight about them is more useful than pretending one dropdown solves everything.
The community's portable persona format. OpenRogue has no card importer and no card library — the working equivalent is a Project whose custom instructions hold the character definition, which then applies to every chat inside it.
Free, open source and endlessly configurable, but it is a client, not a model. It still needs a brain behind it, which is a separate decision from this one and the reason both pages exist.
No model on this list remembers anything outside its context. Long-running characters need you to keep a running summary in the project instructions — an annoyance, not a solved problem.
If you want one answer: start on Euryale 70B. It is the least likely to disappoint on any given evening, its 131K context means you will not hit a wall mid-story, and it is available from the Rogue tier rather than the top of the range. Move to Magnum v4 when you find yourself caring about the sentences, and drop to Rocinante when you find yourself caring about the pace.
The more useful habit is not picking a winner at all. Because you can change models inside a conversation without losing it, the practical workflow is to draft fast on a 12B, escalate to a 70B when a scene turns important, and let the 405B handle the chapter where everything has to land at once. That is the actual argument for a library over a single model, and it is worth more than any ranking.
There is no persona marketplace, no card importer, and no browsing thousands of community characters. Character.AI and the card sites beat OpenRogue outright on discovery, and if that browsing is the appeal, this is the wrong product.
131K is a lot of room but it is a ceiling, not persistence. Nothing here remembers a character across separate conversations without you re-supplying the context.
No avatars, no scene illustration, no image generation of any kind. Platforms that pair roleplay with pictures do something OpenRogue does not do at all.
No refusal layer means the model does not moralise at adults about fiction. It does not mean anything goes: sexual content involving minors is prohibited outright, and no part of this product exists to serve it.
Magnum v4 72B produces the best prose and Euryale 70B is the best all-rounder — if you are unsure, start with Euryale, because its 131K context means a long session will not hit a memory wall. Rocinante 12B is the pick when you want instant replies more than literary ones.
Yes. They are uncensored at the fine-tune level rather than filtered at the platform level, so there is no separate moderation layer to interrupt a scene, and the roleplay tunes were trained to follow the scene rather than break from it. They still need direction, though — a vague character definition produces a vague character regardless of the model.
It depends entirely on the model's context window, which ranges from 8K on Lunaris to 131K on Euryale and the Hermes models. When the window fills, the oldest messages fall out of memory first. The practical workaround is to keep a short running summary of the story in the project instructions so the essentials never age out.
Not as a file — there is no card importer. Paste the character's definition into a Project's custom instructions and every chat in that project will play the persona. The advantage of doing it that way is that the definition is model-independent, so you can swap the model behind the character mid-story.
Uncensored models are ordinary open-weights software without refusal training, and fiction between consenting adults is legal. You remain responsible for what you do with any tool. Content sexualising minors is prohibited and is not something any model on this list is intended for.