The best uncensored LLM GGUF builds: ten you can download or call by API

Ten uncensored and abliterated builds that ship as a GGUF on Hugging Face and answer on one prepaid key. For each: the base, the method, the Q4_K_M file size, today’s rate per million tokens and for 100 messages, and whether it passed today’s real-model check.

Published November 6, 2026 · prices as of November 17, 2026

Most lists of the best uncensored LLM GGUF files sort by what fits a graphics card. This one only includes builds that are both a GGUF on Hugging Face and a model id on one API key, so every row shows the download size next to the price per million tokens. Order follows the leaderboard: today's real-model check first, output price second, both read live. If you test systems for a living and want the same model from a terminal agent without a GPU, the id column is the one to copy.

Ten GGUF builds with a model id on this key, ranked

Each row below lives in two places: a GGUF repo on Hugging Face, checked on October 27, 2026, and a model id you can send to this API. Builds that passed the real-model check today come first; within that group the lower output price per million tokens wins. That is the ordering the leaderboard uses too. Today 5 of 10 passed (run at 2026-11-17 22:18 UTC). Pricing comes from the catalog dated 2026-11-17.

#ModelBuildQ4_K_MOutput $/M100 messagesCheck
1Cydonia 24B v4.3Roleplay14.3 GB$0.60$0.21Steady since Oct 23
2UnslopNemo 12B v4.1Roleplay7.48 GB$1.972$0.85Steady since Nov 14
3Llama 3.3 70B AbliteratedAbliterated42.52 GB$2.80$1.20Steady since Oct 29
4Skyfall 36B v2Roleplay22.4 GB$2.80$0.85Steady since Oct 23
5Qwen3.8 27B ObliteratedAbliterated16.9 GB$7.27$0.45Steady since Oct 22
6Magidonia 24B v4.3Roleplay14.3 GB$0.4235$0.15Changed Nov 14
7Stheno 8B v3.2Roleplay4.92 GB$0.7035$0.30Changed Nov 14
8Anubis 70B v1.1Roleplay42.5 GB$1.085$0.47Changed Nov 15
9Mag Mell R1 12BRoleplay7.48 GB$1.7255$0.74Changed Nov 15
10NeuralDaredevil 8B AbliteratedAbliterated4.92 GB$1.76$0.76Changed Nov 14

Q4_K_M is the file size each repo page printed on October 27, 2026. The 100-message column is 100 × 4,000 input tokens at the input rate plus 100 × 300 output tokens at the output rate, worked out when the page renders.

Top of the table today: Cydonia 24B v4.3 at $0.60 per million output tokens, or $0.21 for 100 messages. The lightest file is Stheno 8B v3.2, 4.92 GB at Q4_K_M.

Base model and method, row by row

Base and method are taken from the Hugging Face card of each build, the way its model page here records them. The right-hand column quotes the GGUF repo.

ModelBase modelMethodRepo note
Cydonia 24B v4.3Mistral Small 3.2 24B Instruct 2506Fine-tune for roleplay and story writingUses the Mistral v7 Tekken template; the card recommends iMatrix quants over these files.
UnslopNemo 12B v4.1Mistral Nemo 12BFine-tune for adventure writing and roleplayAt Q4 this 12B sits inside an 8 GB card.
Llama 3.3 70B AbliteratedLlama 3.3 70B InstructAbliteration, no retrainingQuants of a Llama 3.3 70B Instruct abliteration; the IQ2_M file at 24.12 GB squeezes onto a single 24 GB card.
Skyfall 36B v2Mistral Small 2501 (upscaled)Upscale with continued training for creativity and RPThe card recommends Mistral v7 Tekken, with Alpaca as the alternative for branching stories.
Qwen3.8 27B ObliteratedQwen3.8 27BBlended abliteration (SVD and LEACE)The original repo holds GGUF quants of its own from Q2_K up to Q8_0.
Magidonia 24B v4.3Magistral Small 2509Fine-tune for roleplay and story writingCydonia’s sibling on a Magistral base, same template, different writing style.
Stheno 8B v3.2Llama 3 8BFine-tune, sixth iteration, merged back to its baseTemplate Llama 3 Instruct. Suggested sampling on the card: temperature 1.12 to 1.22, min_p 0.075.
Anubis 70B v1.1Llama 3.3 70B InstructFine-tune for character adherence and unalignmentTemplate Llama 3. At Q4 it needs a pair of 24 GB cards or a Mac with 64 GB.
Mag Mell R1 12BMistral Nemo Base 2407 (ChatML)Multi-stage merge: SLERP specialists, then DARE-TIESTemplate ChatML. The card’s testing used temperature 1.25 with min_p 0.2.
NeuralDaredevil 8B AbliteratedDaredevil 8B Abliterated (Llama 3 8B)Abliteration, then DPOAbliteration followed by a DPO pass to win back quality; laptop-sized.

Cards read October 20, 2026; repo notes read October 27, 2026.

What fits the lab box

For a local run the question is VRAM. At Q4_K_M the two 8B builds (Stheno, NeuralDaredevil) are about 5 GB and run on a laptop. The two 12B builds (UnslopNemo, Mag Mell) are about 7.5 GB and fit an 8 GB card. Cydonia and Magidonia at 14.3 GB want a 16 GB card. Skyfall 36B at 22.4 GB wants 24 GB. The two 70B rows need two cards or a 64 GB Mac. Those figures are what the repos printed on October 27, 2026.

Seven of the ten are chat fine-tunes built for long-form conversation; the Build column says which. For reading a proof of concept, reviewing code or drafting a finding, the three abliterated rows keep a general base model's instruction following and are the better start. The CTF page and uncensored coding AI cover the choice for security tasks in more detail.

Over the API you pay per token instead of per gigabyte of VRAM, which is why the first table prices each build per 100 messages. The docs list the base URL and model id fields for the agents and scripts that use it.

The abliterated rows

Three builds here were abliterated, not fine-tuned: someone found the refusal direction in a released model's weights and removed it, adding no training data. qwen3.8-27b-obliterated was made from Qwen3.8 27B. llama-3.3-70b-abliterated was made from Meta's 70B. neuraldaredevil-8b-abliterated was made from an 8B, then given a DPO pass to repair what the edit cost. Choose one when you want a known base model that answers a question about an in-scope target instead of declining it. Abliterated vs uncensored compares the two methods across this catalog, and what is abliteration explains the edit.

Same weights, two ways to pay

A GGUF is the one-file format llama.cpp loads; Ollama, LM Studio, KoboldCpp and Jan read it too. Hugging Face's GGUF documentation lists the quant types, and puts Q4_K at roughly 4.5 bits per weight. Running locally costs you the hardware plus the time to download, update and swap builds, and nothing leaves the machine. Calling the API costs the per-token rate in the table.

terminal
1# one build, local and by API
2ollama run hf.co/<user>/<model>-GGUF:Q4_K_M
3
4curl https://api.pentestllm.ai/v1/chat/completions \
5 -H "Authorization: Bearer sk-ptl-live-…" \
6 -H "Content-Type: application/json" \
7 -d '{"model": "cydonia-24b-v4.3", "messages": [{"role": "user", "content": "Summarize the open services in this nmap output from an in-scope host"}]}'

Go local when the engagement contract keeps client data on your own hardware, or when a GPU is already sitting idle. Go API when there is no GPU, when you want to try several builds by swapping one id, or when one model has to serve a terminal agent, a script and the playground at the same time. The leaderboard ranks every build on the key and goes further into the trade-off.

Live rates for each id on this page, sortable, from the catalog dated 2026-11-17:

Uncensored models and their prices per million tokens
Magidonia 24B v4.3magidonia-24b-v4.3Roleplay$0.3535$0.4235
Cydonia 24B v4.3cydonia-24b-v4.3Roleplay$0.484$0.60
Stheno 8B v3.2stheno-8b-v3.2Roleplay$0.7035$0.7035
Anubis 70B v1.1anubis-70b-v1.1Roleplay$1.085$1.085
Mag Mell R1 12Bmag-mell-12b-r1Roleplay$1.7255$1.7255
NeuralDaredevil 8B Abliteratedneuraldaredevil-8b-abliteratedAbliterated$1.76$1.76
UnslopNemo 12B v4.1unslopnemo-12b-v4.1Roleplay$1.972$1.972
Skyfall 36B v2skyfall-36b-v2Roleplay$1.925$2.80
Llama 3.3 70B Abliteratedllama-3.3-70b-abliteratedAbliterated$2.80$2.80
Qwen3.8 27B Obliteratedqwen3.8-27b-obliteratedAbliterated$0.582$7.27

Prices per million tokens, read from the live catalog on 2026-11-17.

You prepay from $10 and the balance settles per token.

The ranking rule, and what is missing

Two public inputs decide the order. Every day one fixed prompt goes to each model in the catalog, and the check records whether the input token count reported matches the count that build gave before. A build with a pass ranks above one that failed or was not checked. Inside each group the cheaper output rate per million tokens wins. No quality score is used, because a benchmark result moves with the prompt set and the judge, whereas the rate is what your key is actually charged. The UGI Leaderboard rates how uncensored each model is, and the leaderboard post prints UGI's numbers next to the builds it covers.

Not on this list: builds on the key with no public GGUF repo found on October 27, 2026, which includes the GLM-5.3, DeepSeek V4 and Kimi K3 edits that run by API only, and GGUF repos whose weights are not on the key, Luna AI Llama2 Uncensored, a fixture of the 2023 lists, is one. Each repo on the list served a page on October 27, 2026; one that goes private or disappears is dropped when the list is next reviewed.

Questions

Under this page's rule, Cydonia 24B v4.3: it passed today's check and has the lowest output price among the builds that did, $0.60 per million tokens. Its Q4_K_M GGUF is 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.

The abliterated rows, sized to the card: neuraldaredevil-8b-abliterated on a laptop, qwen3.8-27b-obliterated at 16.9 GB for Q4_K_M, llama-3.3-70b-abliterated when two cards or a 64 GB Mac are available. They keep the base model’s instruction following, which matters for code review and report drafting. All three also answer by API on this key. See the abliterated builds.

Yes, all ten. Send the model id to https://api.pentestllm.ai/v1 with a key from a script, a coding agent or the playground. The GGUF is the same weights if you later want a local copy. See the client setup docs.

Among the three abliterated rows: qwen3.8-27b-obliterated has the widest context and the most downloads on its card, llama-3.3-70b-abliterated has the strongest base, and neuraldaredevil-8b-abliterated runs on almost anything. The leaderboard post orders every abliterated build on the key by today’s check and price. See abliterated vs uncensored.

Two of them. UGI on Hugging Face scores how uncensored a model is. The one on this site ranks the builds you can call by API, using the daily check and output price, and shows UGI’s score next to the models UGI has tested. See the uncensored LLM leaderboard.

Search Hugging Face for the model name with GGUF and choose a quant; Q4_K_M suits most cards. Ollama and LM Studio can pull by name, for instance ollama run hf.co/<user>/<model>-GGUF:Q4_K_M.

A 4-bit K-quant from 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 keeps more quality at about double the size; Q2_K halves the file again and the loss shows.

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.

Get a keyPrepaid from $10, no subscription.

New models and price changes, once a month.

New uncensored builds with their prices, price moves, and what the daily model check found.

Help