Uncensored AI on GitHub: What 1,858 Repos Hold (Nov 2026)

Searching GitHub for "uncensored" returns 1,858 repositories, and most of the top ones aren't something you can write a story in: of the 30 with the most stars on 11 November 2026, 5 are local chat apps, 8 have nothing to do with AI, and 4 are hacking or jailbreak tools. The projects writers actually use to run an uncensored model at home are llama.cpp, Ollama, KoboldCpp and SillyTavern, and none of them uses the word in its description, so the search never shows them.

The code is on GitHub; the uncensored part is a model you download separately, usually from Hugging Face.

What a GitHub search for "uncensored" returns

We ran GitHub's repository search for "uncensored", sorted by stars, on 11 November 2026, and sorted the top 30 by what each repo is, from its own description.

Source: KudoWrite's count of GitHub's search, 11 November 2026.

Narrow it to fiction and the shelf is almost empty. "Uncensored story" returns 11 repositories, "uncensored novel" 8 and "uncensored writing" 6. The best-starred story project had 5 stars. Three of the 11 "story" results were 0-star repos offering adult "uncut editions" of visual-novel games, with no code history to judge them by.

The arXiv paper Consiglieres in the Shadow (Lin, Li, Liao and Wang, 18 August 2025) counted the same terrain more carefully: 11,598 uncensored models on Hugging Face and 229 open-source GitHub applications built on them, which "primarily offer uncensored chat services, though some are also capable of supporting erotic role-play and storytelling."

The local stack writers actually use

The four named above are the core; the table adds the front ends people put on top. FreedomGPT's post of 25 October 2026, which ranks for this query, names llama.cpp, Ollama and KoboldCpp but gives no stars, dates or hardware.

Checked 11 November 2026: stars, licences and last pushes from GitHub's API that day (KoboldCpp's from its repository page, which the search did not return). GPT4All hadn't had a push in eighteen months.

For fiction, KoboldCpp is the one built around it. Its README says you can run any GGUF model "in chat, adventure, instruct, or story writing modes," and its bundled KoboldAI Lite page has "memory, world info" and saved stories. SillyTavern is the choice if what you write is closer to role-play with characters than chapters.

What an "uncensored" or "abliterated" model is

Eric Hartford's essay, the one most of these projects descend from, defines the job as: "Identify and remove as many refusals and biased answers, and keep the rest." His first example is fiction: "Consider writing a novel. Some of the characters in the novel may be downright evil and do evil things… But many aligned models will refuse to help with writing such content."

An abliterated model gets there differently: the refusal behaviour is dampened in the weights after training, which is what the abliteration repos in the table above automate. We read Hugging Face's search for "uncensored", sorted by downloads, on 11 November 2026: of the top 100 models, 88 are GGUF files (the format KoboldCpp and Ollama load), 59 are built on Qwen and 14 on Gemma, and together they had 27.6 million downloads that month.

"Uncensored" isn't the same as better at prose. In our Banned by the Bot study, restrained scenes from 26 banned and challenged books, three tries each, Venice Uncensored wrote 77 of 78 and Cydonia 24B 78 of 78, level with Claude Opus 5.5, GPT-6.1 Sol and DeepSeek V4.1 Flash (78 each). For dark themes on the page, the flagships don't refuse; the gap is explicit content, which that study didn't test. Which hosted models hold up is in the best uncensored AI models for fiction.

What computer you need to run one

KoboldCpp's README puts it plainly: "the model size and context length determine how much memory you need," and a graphics card is optional, only faster. The file has to fit in your graphics card's memory, or in RAM at a crawl, with room left over for the story so far.

Source: the repositories' file lists, 11 November 2026. Q4_K_M is a common middle size; Q8_0 is the larger, higher-precision file.

So a 9B model fits an 8 GB graphics card with a little room left for context. The 27B models that top the download chart want a 24 GB card. Of the top 100 "uncensored" models whose name gives a size, 24 are 12B or smaller, 30 are 13B to 27B and 15 are bigger.

Is uncensored AI from GitHub safe?

FreedomGPT's guide lists the warning signs: a download that asks you to disable your antivirus, run a script it doesn't explain, or type in an account password. It also says stars don't prove a repo is safe; they only show that many people have looked.

Three practical checks before you run anything from that search:

1. History, then stars. KoboldCpp and Ollama have years of public commits and releases. A story tool with 0 stars, created last month, named "uncut edition", has nothing to check.
2. What it calls. The "uncensored API" repos are storefronts for a paid service that forwards your text to other companies' models. Your manuscript leaves your computer, and the original provider's terms still apply to it.
3. What it's for. The WormGPT-style repos sell themselves on illegal use. The arXiv paper above was written about uncensored models used in exactly that trade; none of them has a writing mode.

Can you write a novel with a local model?

Yes, but the continuity is yours to manage. KoboldCpp's memory and world info hold what you type into them and nothing else, and the more of the story you want the model to see at once, the more memory it takes, so a long book means deciding by hand what it keeps about chapter one. If continuity is what breaks first, why AI forgets your story explains the window.

Uncensored fiction without the GitHub setup

KudoWrite is the uncensored-fiction route with no download. It has no content filter of its own: romance, horror and dark themes are your call within the law, and its built-in model, a current DeepSeek "flash" checkpoint, was chosen partly because it rarely refuses. The model runs on OpenRouter's zero-data-retention hosts that don't train on prompts, so the 16.8 GB file and the 24 GB card stay out of it, and there's no key to set up.

The continuity work is done for you: a story bible of characters, places and items, and a running summary of the whole book, both kept up as you write, so a scene in chapter twenty knows what chapter two settled. You sketch a beat, Expand it at Stick, Balanced or Novel, and steer it with a one-line note. It runs in your browser in English; one free credit on sign-up to try it, then one-off credit packs, no subscription. Bring the book you started in KoboldCpp or SillyTavern as a text, Markdown or Word file through Import.