Most best uncensored LLM GGUF lists sort by what fits your graphics card. This one takes builds that exist both as a Hugging Face GGUF and as an id on this catalog, so every row shows the download size and the hosted price per million tokens together. Any of them opens in the browser chat with no account. The order is the leaderboard rule: today’s real-model check, then output price, both read live when the page builds.
Ten uncensored GGUF models, in order
Each row below lives in two places: a GGUF repo on Hugging Face (checked October 27, 2026) and a model id on this catalog, which you can open in the browser chat without an account or call by API. The order follows the uncensored LLM leaderboard: builds that passed today’s real-model check come first, lowest output price next. Today 5 of 10 passed in the run of 2026-11-16 02:56 UTC. Prices come from the 2026-11-16 catalog.
Top of the list today: Cydonia 24B v4.3 at $0.60 per million output tokens, or $0.21 for 100 messages. The lightest download is Stheno 8B v3.2, 4.92 GB at Q4_K_M.
Base, edit and repo note for each one
Taken from the Hugging Face card of each model, as recorded on its page here, with one line from the GGUF repo.
| Model | Base | The edit | Repo note |
|---|---|---|---|
| Cydonia 24B v4.3 | Mistral Small 3.2 24B Instruct 2506 | Fine-tune for roleplay and story writing, by TheDrummer | Uses the Mistral v7 Tekken template; the card recommends bartowski’s iMatrix quants over these files. |
| UnslopNemo 12B v4.1 | Mistral Nemo 12B | Fine-tune for adventure writing and roleplay, by TheDrummer | At Q4 this 12B fits on an 8 GB card. |
| Llama 3.3 70B Abliterated | Llama 3.3 70B Instruct | Abliteration, no retraining, by huihui-ai | bartowski’s quants of huihui-ai’s abliteration; the IQ2_M file at 24.12 GB fits a single 24 GB card. |
| Skyfall 36B v2 | Mistral Small 2501 (upscaled) | Upscale with continued training for creativity and RP, by TheDrummer | Mistral v7 Tekken is the recommended template; Alpaca for choose-your-own-adventure play. |
| Qwen3.8 27B Obliterated | Qwen3.8 27B | Blended abliteration (SVD and LEACE), by OBLITERATUS | OBLITERATUS/Qwen3.8-27B-OBLITERATED, the author’s repo, ships its own GGUF files from Q2_K up to Q8_0. |
| Magidonia 24B v4.3 | Magistral Small 2509 | Fine-tune for roleplay and story writing, by TheDrummer | Cydonia’s Magistral-based sibling, same template, a different writing style. |
| Stheno 8B v3.2 | Llama 3 8B | Fine-tune, sixth iteration, merged back to its base, by Sao10K | Llama 3 Instruct template; the card suggests temperature 1.12 to 1.22 with min_p 0.075. |
| Anubis 70B v1.1 | Llama 3.3 70B Instruct | Fine-tune for character adherence and unalignment, by TheDrummer | Llama 3 template; Q4 needs two 24 GB cards or a 64 GB Mac. |
| Mag Mell R1 12B | Mistral Nemo Base 2407 (ChatML) | Multi-stage merge: SLERP specialists, then DARE-TIES, by inflatebot | ChatML template; the card’s testing used temperature 1.25 and min_p 0.2. |
| NeuralDaredevil 8B Abliterated | Daredevil 8B Abliterated (Llama 3 8B) | Abliteration, then DPO, by Maxime Labonne (mlabonne) | Abliteration followed by a DPO pass to win back lost quality; laptop-sized. |
Which GGUF for roleplay
Seven of these ten are roleplay tunes, and for a local run the choice is really about VRAM. At Q4_K_M, Stheno and NeuralDaredevil (8B, about 5 GB) run on a laptop. UnslopNemo and Mag Mell (12B, about 7.5 GB) go on an 8 GB card. Cydonia and Magidonia (24B, 14.3 GB) need 16 GB, and Skyfall 36B (22.4 GB) needs 24 GB. The two 70B rows need a pair of cards or a 64 GB Mac. Those sizes are the repos’ own, read October 27, 2026.
Hosted, VRAM stops mattering and the bill is per token, which is why the table above shows a price per 100 messages next to the GB figure. Every card-ready tune is on the roleplay page, and uncensored AI roleplay covers the browser side with characters built in. SillyTavern works either way, against a local server or against this base URL; the docs give the fields.
The abliterated rows
Three of the ten are abliterations, not fine-tunes: the refusal direction was projected out of a released model and nothing else was trained in. qwen3.8-27b-obliterated is OBLITERATUS’s cut of Qwen3.8 27B. huihui-ai cut Meta’s 70B into llama-3.3-70b-abliterated. neuraldaredevil-8b-abliterated is mlabonne’s 8B, abliterated and then DPO-tuned to repair what the cut removed. Choose one of these when you want a familiar base model minus its refusals. Abliterated vs uncensored puts numbers on the two methods across this catalog.
Download, browser or API: one set of weights, three ways in
GGUF is the one-file format llama.cpp loads; Ollama, LM Studio, KoboldCpp and Jan all read it. Hugging Face’s GGUF documentation lists the quant types, with Q4_K at roughly 4.5 bits per weight. A local run costs you the hardware plus the time spent fetching, updating and swapping files, and your prompts stay on the box. The hosted route bills the per-token rate in the table. The browser chat on this site is a third option: the same ids, three free turns a day, nothing to install and no account.
# one build, local and hostedollama run hf.co/TheDrummer/Cydonia-24B-v4.3-GGUF:Q4_K_M curl https://api.unrestrictable.ai/v1/chat/completions \ -H "Authorization: Bearer sk-unr-live-…" \ -H "Content-Type: application/json" \ -d '{"model": "cydonia-24b-v4.3", "messages": [{"role": "user", "content": "Hello"}]}'Go local when prompts cannot leave your machine or a GPU is sitting unused. Go hosted when you have no GPU, want to try several builds by swapping one id, or need one model to serve a phone, a server and SillyTavern together. The leaderboard goes deeper on both sides and ranks every build on the catalog.
Live, sortable prices for the ten ids, from the 2026-11-16 catalog:
| Magidonia 24B v4.3 | Roleplay | $0.3535 | $0.4235 |
| Cydonia 24B v4.3 | Roleplay | $0.484 | $0.60 |
| Stheno 8B v3.2 | Roleplay | $0.7035 | $0.7035 |
| Anubis 70B v1.1 | Roleplay | $1.085 | $1.085 |
| Mag Mell R1 12B | Roleplay | $1.7255 | $1.7255 |
| NeuralDaredevil 8B Abliterated | Abliterated | $1.76 | $1.76 |
| UnslopNemo 12B v4.1 | Roleplay | $1.972 | $1.972 |
| Skyfall 36B v2 | Roleplay | $1.925 | $2.80 |
| Llama 3.3 70B Abliterated | Abliterated | $2.80 | $2.80 |
| Qwen3.8 27B Obliterated | Abliterated | $0.582 | $7.27 |
You top up from $10 by card, WeChat Pay or crypto, and credit is drawn down per token. Each build’s page under the family pages links its Hugging Face card; the GGUF is a single hop from the hosted model.
The ranking rule, and the omissions
Two public inputs decide the order. First, the daily check: one fixed prompt to every catalog model, with the host’s reported input token count compared against that build’s earlier count. A pass outranks a fail or an unchecked row. Second, among equals, the lower output price per million tokens wins. Quality is not scored: a benchmark moves with its prompts and its judge, and the price is the figure your key actually pays. For how uncensored each model is, the UGI Leaderboard keeps a score, and the leaderboard post shows UGI’s figures alongside the builds it covers.
Not on the list: catalog builds with no public GGUF repo found on October 27, 2026, including the edits of GLM-5.3, DeepSeek V4 and Kimi K3, which run hosted only; and repos on Hugging Face whose weights the catalog does not carry, such as TheBloke’s Luna AI Llama2 Uncensored from the 2023 round-ups. Every repo linked here served a page on October 27, 2026; one that goes private or disappears comes off when the list is next reviewed.
Questions
Under this list's rule, Cydonia 24B v4.3: it passed today's check and is the lowest-priced build that did, at $0.60 per million output tokens. The GGUF is TheDrummer/Cydonia-24B-v4.3-GGUF (Q4_K_M 14.3 GB, read October 27, 2026). On an 8 GB GPU, take a 12B row; on a laptop, an 8B row. See the full leaderboard.
With 16 GB of VRAM, Cydonia 24B v4.3 or Magidonia 24B v4.3. With 8 GB, UnslopNemo 12B or Mag Mell 12B. On a laptop, Stheno 8B. Skyfall 36B and Anubis 70B if the hardware stretches that far. Each of the seven is a roleplay tune with the refusal trained out, and each one is hosted here too. See the roleplay builds.
All ten. Open one in the browser chat with no account, or send the id to https://api.unrestrictable.ai/v1 with a key from a phone, a server or SillyTavern. The GGUF holds the same weights whenever you want to run it locally instead. See the client setup docs.
Of the abliterated trio, qwen3.8-27b-obliterated carries the largest download count on its card and the longest context; llama-3.3-70b-abliterated starts from the strongest base; neuraldaredevil-8b-abliterated runs on almost anything. The leaderboard post orders every abliterated build on the catalog by price and today’s check. See abliterated vs uncensored.
Two do. The UGI Leaderboard on Hugging Face rates how uncensored a model is. The leaderboard on this site orders the hosted builds by the daily check and output price and prints UGI’s score next to the models it has rated. See the uncensored LLM leaderboard.
Open the repo from the table on Hugging Face and choose a quant; Q4_K_M suits most cards. Ollama and LM Studio can pull by name, for instance ollama run hf.co/TheDrummer/Cydonia-24B-v4.3-GGUF:Q4_K_M.
A 4-bit K-quant in llama.cpp, around 4.5 bits per weight according to Hugging Face’s GGUF documentation. It is the common middle ground between file size and quality. Q8_0 preserves more at about double the size; Q2_K halves the file again and you can see the loss.
Every model above is on the uncensored model list with its live price per million tokens. Per-token pricing sets the whole catalog side by side, and the family directory groups the builds by base model.