Ranked on prose, not on parameter count.
Writers hit two different walls with mainstream assistants, and they are usually mistaken for one wall. The first is the refusal: the antagonist gets softened, the dark chapter gets reframed as a lesson, the scene fades out. The second is the prose itself — the shivers down spines, the testaments to, the voices barely above whispers, an accumulated house style you can smell from across the room.
Removing the filter fixes the first wall and does nothing at all about the second. Plenty of uncensored models write badly. So this ranking weighs prose quality and cliché load as heavily as willingness, which produces a different order than a general capability list would.
Every model here runs on OpenRogue, and because you can change models inside one conversation, the ranking matters less than it looks. Most serious drafting sessions end up using two or three of these for different jobs.
Rhythm, restraint, and the willingness to write a short sentence. This is the criterion most rankings skip because it is hard to measure and it is also the one writers actually feel.
How much of the output is recycled AI phrasing you will have to sand off before a human reads it. One model on this list was fine-tuned specifically against that, and it shows.
A model that will not write the ending the story has been building toward is not a writing tool. Every entry here commits to the scene, including the uncomfortable ones.
Holding a plot, a cast and a voice across tens of thousands of words. This is where context length stops being a spec sheet number and starts being the difference between a chapter and a mess.
Give it a voice — yours, or an author's — and see whether it holds the register for pages or drifts back to its defaults by the third paragraph.
This ranks models for prose and long-form work. Interactive fiction where you play a character in real time is a different job with a different winner, and it has its own page. Non-fiction that mainly needs correctness rather than voice is better served by the general ranking, and code is on the coding page.
| Rank and name | Size | Context | Best at | Plan |
|---|---|---|---|---|
| 1. Magnum v4 | 72B | 32K | Sentence-level prose, style transfer | Pro |
| 2. UnslopNemo | 12B | 32K | Human-sounding drafts, de-slopping | Rogue |
| 3. Hermes 3 405B | 405B | 131K | Structure, multi-POV plotting | Pro |
| 4. Euryale | 70B | 131K | Continuation, long serials | Rogue |
| 5. Hermes 4 | 70B | 131K | Outlining, revision notes, plot logic | Rogue (Thinking mode on Pro) |
| 6. WizardLM-2 | 141B MoE | 65K | Structured documents, multi-constraint briefs | Rogue |
| 7. Frontier models | Vendor-hosted | Large | Polish, non-fiction, correctness | Pro |
The best prose engine here, and the one that will write the ending.
Anthracite trained Magnum v4 toward one target: the prose quality of the frontier Claude models, without inheriting their refusals. That is a narrow and unusual goal for an open fine-tune — most optimise for compliance or for benchmark scores — and it produces a model that writes with restraint. It will use a short sentence. It will leave something unsaid.
For fiction it is the obvious first choice. Feed it a passage and ask for continuation in the same voice and it holds the voice; give it a premise other models would soften and it goes where the premise leads. The catch is structural rather than stylistic, and it is the 32K context.
Best at:
Worse at:
Pick this if: Scenes and chapters where the sentences are the product. · Full Magnum v4 page
The lowest cliché load in the library, at drafting speed.
'Slop' is the community's word for the recycled phrasing every model overuses, and TheDrummer built UnslopNemo as a direct attack on it: a Mistral Nemo 12B tune trained on data curated to break those patterns rather than reinforce them. The result is not a smarter model. It is a model whose first draft needs less repair.
That makes it the right pick for anything going somewhere humans will read it unedited — newsletter copy, blog drafts, dialogue passages. It is also the best de-slopping tool on the list: hand it a paragraph another model wrote and ask for the AI-isms removed, and it knows what you mean.
Best at:
Worse at:
Pick this if: First drafts that a person will read without you rewriting them first. · Full UnslopNemo page
The architect. Use it on the structure, not the sentences.
Nous Research's fine-tune of Llama 3.1 405B is the largest genuinely uncensored model available online, and its advantage over the 70B class is not prettier sentences — it is holding more of a work in mind at once. Seven point-of-view characters, three timelines, a reveal that has to be seeded four chapters early: that is where 405B separates from everything below it.
The practical pattern is to use it for the parts of writing that are not writing. Full-series synopses, chapter breakdowns, continuity checks against 131K of manuscript, honest structural criticism of a draft. Then draft the actual prose on Magnum and bring it back.
Best at:
Worse at:
Pick this if: Outlining, continuity and structural criticism on long works. · Full Hermes 3 405B page
The chapter-continuation model — long memory, reliable voice.
Sao10K's Llama 3.3 tune is best known for roleplay, but the property that earns it a place here is the same one: it writes expressively without losing track of what it already wrote. With 131K of context you can paste the preceding chapters, ask for continuation, and get something that remembers the established voice rather than inventing a new one.
It is the pragmatic middle of this list. Not the most beautiful prose and not the deepest structural thinker, but the model most likely to produce usable pages on demand from a Rogue-tier subscription.
Best at:
Worse at:
Pick this if: Continuing a long work in its established voice. · Full Euryale page
The editor, not the writer. Best paired with Thinking mode.
Hermes 4 is a hybrid reasoner: it can work step by step when asked and answer directly when not. For writing that is less useful for producing prose than for interrogating it — plot-hole hunting, motivation checks, arguing about whether a reveal has been earned, and producing revision notes that are specific instead of encouraging.
Nous trains the Hermes line for neutral alignment, which for an editorial model means it will actually tell you a chapter is not working. Paired with deep-reasoning mode on Pro it reads a draft as a structural argument rather than a piece of text to compliment.
Best at:
Worse at:
Pick this if: Revision passes and structural editing on a draft you already have. · Full Hermes 4 page
For long structured non-fiction that has to hold its shape.
The Mixtral-based mixture-of-experts that Microsoft published and then withdrew is the specialist here, and its specialism is instruction density. Long briefs with a dozen constraints — sections, word counts, tone shifts, a table halfway through — are exactly what it was benchmarked on, and it holds them where smaller models start dropping requirements silently.
For fiction it is outclassed by the models above. For a technical guide, a long report, a specification or anything with a required shape, it is the most reliable model in the library at producing the shape you asked for.
Best at:
Worse at:
Pick this if: Long non-fiction with a required structure. · Full WizardLM-2 page
The best line-level polish available — and the filtered option.
The Pro tier includes frontier models from the major vendors alongside the uncensored library, and it would be dishonest to leave them off a writing ranking. On clean non-fiction, line editing and anything requiring factual care, they remain the strongest tools in the dropdown.
They are ranked last on this page because this page is about uncensored writing, and they are the models that will decline the dark chapter. The realistic workflow is to use them for the manuscript's safe two-thirds and switch to Magnum for the third they will not touch — which is a switch you can make inside one conversation.
Best at:
Worse at:
Pick this if: The parts of a manuscript that never collide with a content policy.
The single best uncensored writing model is Magnum v4 72B, and it is not especially close on prose. If you write fiction and care about sentences, that is the answer and the rest of this page is about the jobs it is not best at.
The better answer is a rotation. Structure and continuity on Hermes 3 405B, drafting on Magnum, de-slopping and fast iteration on UnslopNemo, revision notes on Hermes 4, continuation of a long work on Euryale. Switching models inside a conversation is what makes that practical, and it is worth more to a working writer than any single model on the list.
No document canvas, no track changes, no scene reordering, no chapter tree. You write in a chat window and move text out yourself. Dedicated AI writing suites beat this outright on workflow.
There is no structured world bible that gets injected automatically when a name comes up. Project custom instructions are the closest equivalent and they are a blunter tool.
Magnum v4 carries 32K. For novel-length work that is a genuine constraint, and the workaround — summarise, paste, continue — is manual labour.
Removing a refusal layer does not remove hallucination. Anything factual still needs checking, and a model willing to answer confidently is not the same as a model that is right.
Magnum v4 72B. Anthracite trained it specifically toward frontier-Claude prose quality without the refusals, and it writes with more restraint and rhythm than anything else in the library. Its limitation is a 32K context window, so very long works need structural help from a 131K model like Hermes 3 405B.
Yes, for adults. These models were trained without the refusal layer, so a dark premise is followed rather than reframed, and a scene ends where the story ends rather than fading out. Content sexualising minors is prohibited regardless of model.
Use UnslopNemo 12B, which was fine-tuned specifically against the recycled phrasing that marks machine-written text, and give whichever model you use an explicit style instruction — naming a register or an author works far better than asking for 'good writing'. Running a finished draft back through UnslopNemo with an instruction to remove AI-isms is a genuinely effective final pass.
Not in one pass, and no honest tool claims otherwise. Context windows run from 32K to 131K, which covers chapters and long sections rather than complete manuscripts. The workable pattern is outlining on a large-context model, drafting chapter by chapter on a prose model, and keeping a running summary so continuity survives between sessions.
For material Claude will not write, Magnum is the only one of the two that is an option at all. For clean non-fiction and line-level polish, the frontier models are still stronger. That is why both sit in the same dropdown — the realistic answer is that a working writer uses both.