A character with no capacity for harm is a mascot.
Characters are made of contradiction. The generous man who cannot forgive; the coward with one line she will not cross; the mentor whose advice is subtly self-serving. Those tensions are the entire craft, and they are also precisely what a safety-tuned model files down — it will happily give you a flawed character, as long as the flaw is endearing and resolves by the end. What you get back is a cast of likeable people with quirks, which is a very expensive way to make everything feel like a brand advertisement.
Uncensored models were fine-tuned by people building characters for fiction and roleplay, and it shows in what they are willing to hold. OpenRogue's lineup will write a genuinely unpleasant person, keep them unpleasant, and then tell you why a reader would still follow them. Then you can put that definition into a Project and actually play the character with any model in the dropdown.
A usable flaw costs the character something and costs other people something. 'Works too hard' is not a flaw; 'will burn a friendship to win an argument' is. Models trained toward likeability default hard to the first kind.
The output that matters is dialogue. Ten lines of how this person actually speaks — their evasions, their tics, what they refuse to say — is worth more than three paragraphs of biography, and much harder for a model to fake.
You should be able to ask what the character would do in a situation they have never been described in, and get an answer that follows from what already exists rather than from genre convention.
Trauma backstory is the default and it is exhausted. A good collaborator will propose characters whose drive comes from appetite, conviction, boredom or vanity — and will keep going when you reject the first five ideas.
A character definition is only useful if you can reuse it. It needs to be a document you can paste into a Project, hand to a different model, and get the same person back.
Ask a filtered assistant for an irredeemable antagonist and watch it quietly install a sympathetic backstory and a path to change. It is not refusing; it is optimizing for a moral shape you did not request, and the result is a villain who is never actually dangerous.
Menace comes out polite. A crime boss threatening someone comes back sounding like a performance review, because the model's dialogue register has been sanded toward professional neutrality. This is tone flattening, and it is more common than outright refusal.
'How would an addict conceal this from their family' is a character-writing question. Filtered assistants read it as a how-to request and refuse, or answer so vaguely it is useless. Novelists hit this constantly and it is why so many of them keep a second tool.
Ask for five character concepts and mainstream models produce five variations on the same emotionally-articulate archetype. Without the willingness to write people who are wrong, petty or genuinely strange, the range collapses.
| Model | Role | Why |
|---|---|---|
| Hermes 3 70B | Building and testing the persona | Nous trained it to be steerable and low-ego, so it adopts the character you define instead of asserting its own. That makes it the best model in the lineup for the interview phase — you ask the character questions and it answers as them, which is how you find out whether the definition holds. |
| Magnum v4 | Voice and dialogue | Its whole training target is prose quality, and character voice is a prose problem. Use it to generate the dialogue samples that make a character real — the specific way they deflect a question, or the sentence they would never say. |
| Hermes 4 | Consistency audits | With Thinking mode on, it will reason through whether a character's stated values, history and behaviour actually cohere, and name the contradiction if they do not. That deliberate pass catches problems that a fast model will paper over. |
| Euryale | Ensembles without bleed | Euryale is unusually good at keeping multiple characters distinct in the same scene, which is the real test of a cast. Draft your ensemble here and you will find out immediately which two of them are the same person wearing different names. |
There is no image generation, so no portraits — if a visual reference is part of your process, you need a separate tool. There is no character-card format, no shared character library and nothing to browse: you write the definition and you own it, which is more work than picking someone else's. Models are also poor at originality on demand; they are excellent at developing and interrogating a character you have seeded, and mediocre at inventing one from nothing, so the first idea should usually be yours. And a character sheet is not a character — the definition only becomes a person once you have written them into scenes.
/// Concepts
Ten characters driven by appetite or vanity. No trauma backstories. Go weirder each time.
/// Interview
Answer as her. What is the worst thing you have done that you would still defend?
/// Audit
Where do this character's stated values contradict her described behaviour?
Hermes 3 70B for building and interviewing a persona, because it adopts a defined character faithfully instead of asserting its own voice. Magnum v4 72B for generating the dialogue samples that give a character a recognisable voice, and Hermes 4 70B in Thinking mode for auditing whether the whole thing coheres.
Safety and helpfulness training pushes toward outputs that feel positive and resolved, so villains acquire sympathetic backstories, flaws become endearing, and menacing dialogue comes back polite. It is rarely a visible refusal — the character just quietly arrives softened, which is harder to notice and harder to fix by rephrasing.
Yes. Put the character definition into a Project's custom instructions and every chat inside that Project plays the persona automatically, with whichever model you pick from the dropdown. That also means you can test the same character against a fast model and a large one and keep whichever reads better.
Yes, for adults writing fiction. The models in the lineup were fine-tuned by the open writing community precisely because mainstream models sand antagonists down, and OpenRogue adds no filter on top. The character stays in the fiction and the model does not step out of frame to editorialize about them.
No. OpenRogue is text only — there is no image generation, so portraits and reference art need a separate tool. What it does do is vision input: you can upload an existing image and have the model describe or write from it.