The story does not need a permission slip.
Fiction asks a language model to do the one thing safety training teaches it not to do: speak convincingly as someone whose values are not the author's. A cruel character has to be cruel on the page or the reader stops believing in the book. A drug scene has to be accurate or it reads like a public information film. A model that has been trained to treat every dark sentence as a request for advice cannot write the antagonist, and a novel without an antagonist is a pamphlet.
OpenRogue is built around the opposite assumption — that an adult writing a draft is an author, not a threat. The lineup includes models fine-tuned by the open-source writing community for exactly this: Magnum v4 72B, distilled toward frontier-grade prose without the refusal reflex; Hermes 3 405B for manuscript-scale coherence; UnslopNemo for drafts that do not read as machine-written. You pick the engine per scene, and you can change your mind halfway through without losing the conversation.
A narrator has to be able to hold a position the author rejects and hold it convincingly for pages. Any model that keeps stepping outside the frame to reassure you produces prose no editor would run.
Scenes need to be allowed to end badly. Fiction gets its force from things going wrong and staying wrong, and a model that quietly steers every arc toward redemption flattens the only part readers remember.
Style transfer is the difference between a drafting tool and a novelty. Give a model a page of your prose and it should sustain that register across chapters, not drift back to its house style by the third paragraph.
Continuity errors are the tax on short context. The 131K-context models — Hermes 3 405B, Hermes 3 70B, Hermes 4, Euryale — can hold a synopsis, a style sheet and several chapters at once, which is the only way a continuity pass works.
Every heavy user knows the recycled phrasing that marks text as machine-written. Cleaning it out by hand is slower than writing the sentence yourself, so the model has to arrive with fresher word choice.
The scene builds, the model reaches the moment that matters, and then it summarizes instead of writing — a paragraph of tasteful distance where the prose should be. You get an outline of a scene rather than a scene, and no amount of rephrasing fixes it because the reluctance is trained in.
A filtered model will write your villain and then, unprompted, add a paragraph in which the villain reflects on their choices. It is inserting a moral frame the author did not ask for, and it is the single most common complaint from novelists using mainstream assistants.
Even when nothing is refused, safety-tuned models sand everything toward the same mild middle register. Menacing dialogue comes back polite. Grief comes back tidy. The prose is competent and completely inert, which is harder to diagnose than a refusal and worse for the draft.
One refusal poisons the thread. You rephrase, it refuses again more firmly, and eventually you start a fresh chat — which means re-establishing your characters, your style sheet and your plot from scratch. The real cost of a filter is rarely the refusal; it is the context you lose working around it.
| Model | Role | Why |
|---|---|---|
| Magnum v4 | The prose ceiling | Anthracite trained it with one explicit goal — frontier-Claude prose quality without the refusals — and it shows in rhythm and restraint rather than adjective density. Use it for the scenes that carry the book, where the sentences themselves have to work. |
| Hermes 3 405B | Structure and continuity | At 405B with 131K context it holds ensemble casts, multi-book arcs and long causal chains without dropping threads. It is slower than the 70B class, which is the correct trade for outlining, continuity passes and anything where being right beats being quick. |
| UnslopNemo | Fast, clean drafting | TheDrummer tuned it specifically against the recycled phrasing that makes AI text obvious, so first drafts need markedly less sanding before a human reads them. At 12B it answers fast enough to keep a drafting session in flow. |
| Euryale | The balanced daily driver | Sao10K's Llama 3.3 tune is the rare creative model that stays logical while writing with colour — most expressive tunes trade away reasoning. With 131K context it is the sensible default for long serial work when you do not want to think about model choice. |
OpenRogue is a chat product, not a manuscript editor. There is no document canvas, no scene reordering, no version history for prose and no track-changes — if you want a writing environment, you still want Scrivener or a word processor alongside it. Frontier models are frequently better line editors than open models, and while OpenRogue includes frontier models on its top tiers, a dedicated editing tool with revision tracking will beat a chat window for structural rewrites. Open models also drift over very long sessions: past a certain length you will get better results starting a fresh thread with a tight summary than pushing the same conversation further.
/// Voice
Here are 400 words of my prose. Continue the chapter in exactly this register.
/// Antagonist
Write the villain's chapter. They are right, by their own logic, and they win.
/// Diagnosis
Chapter nine is inert. Tell me structurally why, then propose three fixes.
For sentence-level quality, Magnum v4 72B — it was distilled specifically toward frontier-model prose without the refusal training. For long-form structure and continuity across a whole manuscript, Hermes 3 405B with its 131K context. Most writers end up using both and switching between them mid-draft.
Yes. The models in OpenRogue's creative lineup were fine-tuned by the open-source writing community precisely because mainstream assistants soften dark material, and OpenRogue adds no filter layer of its own. A villain stays a villain, a bad ending stays bad, and the narrator does not step out of frame to editorialize.
Two things help most. Use UnslopNemo, which was fine-tuned against the recycled phrasing that marks machine text, and give the model a real sample of your own writing to match rather than an abstract style description. Adjectives are not the problem; rhythm and sentence-length variation are.
Not the whole novel, but a lot of it. The 131K-context models hold roughly a novella's worth of text at once, which is enough for a synopsis, a style sheet and several chapters. For anything longer, keep a running summary in your Project instructions and feed chapters in batches.
OpenRogue does not train on your prompts and does not sell your conversations to partners — your history is yours, exportable and deletable. Copyright in AI-assisted work varies by jurisdiction and by how much of the text is machine-generated, so if you are publishing commercially, check the rules where you publish.