Plenty of testers keep an Ollama uncensored model on the laptop for the questions a stock model declines: llama2-uncensored, one of the Dolphin tags, wizardlm-uncensored or an abliterated DeepSeek distill, each a GGUF running on your own GPU. This page lists every official one from the Ollama library as it stood on 27 October 2026, with its download size, and sets the same or newer build on this key beside it with the live per-million price and the cost of 100 messages. It ends with the two-field change that points Open WebUI or AnythingLLM at the key instead of localhost.
If abliterated and uncensored sound like the same thing to you, abliterated vs uncensored separates them in a page.
What the Ollama library offers under "uncensored"
Searching the Ollama library for "uncensored" on 27 October 2026 returned eight official tags plus roughly a dozen under personal namespaces. The official ones fill the table. The tag links to its library page, and the size column repeats what that page shows. Every one of them shows a last push one or two years back.
| Library tag | Download size | Page details |
|---|---|---|
llama2-uncensored | 7B at 3.8 GB; 70B at 39 GB | Llama 2 base; 2.7M pulls; last push 2 years back |
dolphin-llama3 | 8B at 4.7 GB; 70B at 40 GB; 8K context | Dolphin 2.9 over Llama 3; 2.2M pulls; last push 2 years back |
dolphin-mistral | 7B at 4.1 GB; 32K context | Dolphin 2.8 over Mistral 7B; 1.8M pulls; last push 2 years back |
dolphin3 | 8B at 4.9 GB; 128K context | Dolphin 3.0 over Llama 3.1 8B; 4.2M pulls; last push 1 year back |
wizardlm-uncensored | 13B at 7.4 GB; 4K context | 666.5K pulls; last push 2 years back |
The search also surfaced three official tags that are left out because nothing on this key maps to them: dolphin-mixtral at 8x7B and 8x22B, wizard-vicuna-uncensored from 7B to 30B, and everythinglm at 13B with a 16K context. The personal namespaces held mostly abliterated Qwen 3.8 and Gemma 4 conversions, in sizes such as 9B and 27B.
ollama run llama2-uncensored# or any tag in the table: ollama run dolphin-llama3:70bSame families, newer weights, no GPU
For every tag above there is a build on this key from the same family or a later generation, served over the API and billed per token. Pulling the 70B Llama 2 tag means 39 GB of weights on disk; llama-3.3-70b-abliterated is a 70B Llama 3.3 with the refusal direction stripped out. The Dolphin tags stop at 7B and 8B; dolphin-mistral-24b-venice puts Dolphin on Mistral Small 24B. For the R1 distills, deepseek-v4.1-flash-abliterated is a current DeepSeek that was abliterated rather than retrained.
| Library tag | API id | Input per M | Output per M | 100 messages |
|---|---|---|---|---|
llama2-uncensored | llama-3.3-70b-abliterated | $2.80 | $2.80 | $0.64 |
dolphin-llama3 | llama-3.3-70b-abliterated | $2.80 | $2.80 | $0.64 |
dolphin-mistral | dolphin-mistral-24b-venice | $0.50 | $2.25 | $0.17 |
dolphin3 | dolphin-mistral-24b-venice | $0.50 | $2.25 | $0.17 |
wizardlm-uncensored | qwen3.8-27b-uncensored | $1.2125 | $3.6375 | $0.35 |
The daily check column in the next table is read from the proof page: one fixed prompt to every model every day, recording whether the reply still looks like the model named on the label. A local pull offers no such check; a GGUF on disk is whatever it is.
| Wire formats | |||||
|---|---|---|---|---|---|
| Llama 3.3 70B Abliterated | Abliterated | $2.80 | $2.80 | 33K | |
| Dolphin Mistral 24B Venice Edition | Uncensored | $0.50 | $2.25 | 33K | |
| Qwen3.8 27B Uncensored | Uncensored | $1.2125 | $3.6375 | 33K |
| API id | Daily check | Window |
|---|---|---|
llama-3.3-70b-abliterated | Steady since Oct 29 | 32,768 tokens |
dolphin-mistral-24b-venice | Steady since Oct 23 | 32,768 tokens |
qwen3.8-27b-uncensored | Steady since Oct 29 | 32,768 tokens |
Family pages hold the rest of each line: the Llama family page for the Llama rows, the DeepSeek family page for DeepSeek, and the Qwen family page for Qwen. The security-tuned builds, which these general tags never had an equivalent for, are on the models page.
"Ollama deepseek abliterated": the distills versus the API build
When someone searches for an Ollama DeepSeek abliterated model they land on a community upload of the R1 distills. On 27 October 2026 that page offered seven sizes between 1.5B and 70B, every one with a 128K context, 748.7K pulls and a last push one year back. The 32B quant weighs 20 GB and the 70B 43 GB, which is why laptops end up on the 7B, 8B or 14B distills.
The DeepSeek line on this key is deepseek-v4.1-flash-abliterated, the present V4.1 Flash with the refusal direction removed. A quicker lane serves the same weights as deepseek-v4.1-flash-abliterated-fast, so there is no distillation in either one. Moving a request over is a one-host edit from Ollama’s own OpenAI-compatible endpoint, which its docs place at http://localhost:11434/v1:
# Ollama's own OpenAI-compatible endpoint, from its docscurl http://localhost:11434/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{"model": "dolphin-llama3:8b", "messages": [{"role": "user", "content": "Explain what this nmap flag does: -sV --script vuln"}]}'# The same request against this key: change the host and the model idcurl https://api.pentestllm.ai/v1/chat/completions \ -H "Authorization: Bearer sk-ptl-live-…" \ -H "Content-Type: application/json" \ -d '{"model": "llama-3.3-70b-abliterated", "messages": [{"role": "user", "content": "Explain what this nmap flag does: -sV --script vuln"}]}'The second call bills per token at the live rate, with nothing to download, no VRAM budget and no quant to pick. Keys start with sk-ptl-, come from the console and are displayed once. To try the build before making a key, the playground runs it in the browser.
llama2-uncensored and dolphin-llama3: when to stop pulling them
dolphin-llama3 is Dolphin 2.9 on Llama 3 in 8B and 70B with an 8K window. llama2-uncensored is a generation older again: Llama 2 at 7B and 70B with 4K of context. Neither page has moved in two years. They still run, and a 7B on a laptop is a fine way to triage a single log file offline.
At 70B the arithmetic changes. The GGUF is 39 to 40 GB before it loads into memory. llama-3.3-70b-abliterated is Llama 3.3 70B Instruct minus its refusal direction, with the window shown in the table and a per-token price on the same row. A hundred messages of finding review on it cost the figure in the last column; no power bill, no download.
If it is Dolphin you want, dolphin-mistral-24b-venice is the Dolphin tune of Mistral Small 24B, with over three times the parameters of the 7B tag; its window is in the table as well.
Move Open WebUI or AnythingLLM from Ollama to the key
Both front-ends reach Ollama over an OpenAI-compatible connection, and both accept another connection of the same type. Only the URL and the key differ. The picker then loads from /v1/models with every uncensored and security id the key may call.
Open WebUI
Go to Settings > Admin > Connections and add an OpenAI API connection as its connection doc describes. Keeping the Ollama connection alongside puts local and API builds in one picker.
URL https://api.pentestllm.ai/v1API Key sk-ptl-live-…The full treatment, including a picker trimmed on the key and a preset per build, is in the Open WebUI guide.
AnythingLLM
In Settings, change the LLM provider from Ollama to the entry AnythingLLM’s docs name OpenAI (generic), the one meant for developers, then enter the base URL, the key and a model id:
Base URL https://api.pentestllm.ai/v1API key sk-ptl-live-…Model llama-3.3-70b-abliterated- Base URL
- https://api.pentestllm.ai/v1Includes /v1, nothing after it
- Key
- sk-ptl-live-…Goes out as a Bearer token
- Model
- llama-3.3-70b-abliteratedAny id that /v1/models returns
Before either app, list the models from a shell. The call is free and returns the exact ids the picker is about to display.
curl https://api.pentestllm.ai/v1/models -H "Authorization: Bearer sk-ptl-live-…"Any other OpenAI-compatible client, from coding agents to a desktop chat app, takes the same base URL and key; the docs page lists the fields for each one, and the Kali MCP and PentestGPT pages cover the security tools.
Questions
Because the tag Ollama is serving is a stock instruct model. The library builds of llama3, mistral, qwen and gemma keep their alignment, and AnythingLLM forwards your prompt to whatever tag is selected. Either pull one of the uncensored tags in the first table and select it, or switch the provider to this API and choose an abliterated or security id. Those have the refusals taken out of the weights, so no system-prompt workaround is needed. See AI with no restrictions.
On a laptop, the 7B and 8B tags in the first table fit; dolphin3 carries the largest context at 128K. A 70B means a 39 to 43 GB download, and the API rows above serve that size class per token with no hardware. For the ranked view across every build on this key, see the leaderboard. See the uncensored LLM leaderboard.
Only community tags: the search on 27 October 2026 showed abliterated Qwen 3.8 builds in 9B and 27B, and no official entry. On this key qwen3.8-27b-uncensored is the general Qwen build and the rest of the Qwen line is on the uncensored Qwen page. The acceptable use page states what is never permitted. See the uncensored Qwen page.
Yes. On 27 October 2026 the tag was live with 2.7M pulls, 7B and 70B variants and a last push two years earlier. Being Llama 2, it has a 4K context; llama-3.3-70b-abliterated is the present-day Llama at that size on this key.
Not quite. It is the R1 distills, Qwen and Llama models trained on R1 output and then abliterated, from 1.5B to 70B; the 671B DeepSeek weights will not fit a consumer GPU. deepseek-v4.1-flash-abliterated on this key is a true DeepSeek build, and the daily check records whether it keeps answering like one. See the daily real-model check.
Yes. Open WebUI holds several connections side by side. Leave the Ollama connection in place for local tags and add an OpenAI API connection to https://api.pentestllm.ai/v1 with your key; the picker merges both sets of ids.
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.