Run uncensored models with Ollama.

What you'll have working

By the end you'll have:

• an uncensored model chatting entirely on your own computer
• your own saved version of it, with a system prompt and settings you choose
• a way to load almost any uncensored GGUF from Hugging Face

The worked example uses Dolphin 3.0 Llama 3.1 8B, because it runs on the widest range of hardware. The same steps work for every model on the best uncensored models page.

Before you begin

You also need free disk space for the download and a terminal: PowerShell on Windows, Terminal on macOS or Linux. Download sizes are from each model's listing. Memory needs are estimates and grow with context length.

Step 1: Install Ollama

• Windows and macOS: download the installer from ollama.com/download and run it. Ollama keeps running in the background.
• Linux: run the official install script:

curl -fsSL https://ollama.com/install.sh | sh

To check it worked, open a new terminal and run ollama --version. It should print a version number.

Step 2: Run your first uncensored model

ollama run dolphin3

The first run downloads the model. After that it starts in seconds. When you see the >>> prompt, type a message and press Enter.

Check it works: ask something, then a follow-up that depends on the first answer, such as "now rewrite that from the villain's point of view." If the reply uses the earlier context, your setup is working.

Type /bye to leave. The model stays downloaded.

Step 3: Set your own system prompt

Dolphin's maintainers are explicit that the system prompt is how you set the model's tone and rules. Inside a chat, try:

/set system "You are a blunt, uncensored fiction editor. Answer directly. No disclaimers."

Then make it permanent by saving a named copy:

/save my-dolphin

Next time, run ollama run my-dolphin.

Step 4: Build a custom model with a Modelfile

To control temperature, context length and the system prompt in one place, create a plain text file called Modelfile:

FROM dolphin3
SYSTEM """You are a direct, uncensored assistant. Answer the question asked. Skip moralizing and disclaimers unless the user asks."""
PARAMETER temperature 0.7
PARAMETER num_ctx 8192

Then build and run it:

ollama create my-uncensored -f Modelfile
ollama run my-uncensored

num_ctx sets how much conversation the model can remember. Raise it for long chats or documents, but higher values need more memory. For Mistral-24B-based Dolphin models, the maintainers recommend a much lower temperature, around 0.05 to 0.1.

Step 5: Run bigger uncensored models from Hugging Face

Ollama can pull GGUF files straight from Hugging Face using hf.co/<user>/<repo>:<quant>:

ollama run hf.co/lhca521/Huihui-Qwen3.6-27B-abliterated-GGUF:Q4_K_M

Pick a quantization from the repo's file list that fits your memory, using the VRAM guide. Some models publish an Ollama tag directly. For example, the Qwen 3.6 abliterated card gives:

ollama run huihui_ai/qwen3.6-abliterated:27b

New models need a new Ollama. Qwen 3.8-27B builds need a llama.cpp from May 2026 or later. If a fresh model won't load, update Ollama first.

Everyday use

• See what's downloaded: ollama list
• See what's loaded in memory: ollama ps
• Delete a model: ollama rm <name>
• Use a chat window instead of the terminal: Ollama serves a local API at http://localhost:11434, which chat apps and agents on the same machine can connect to. Everything stays local unless you point that app somewhere else.

Troubleshooting

Sources and files

• Ollama download
• Ollama: dolphin3 tags (4.9GB Q4_K_M, 128K context, Llama 3.1 license)
• Huihui Qwen3.6-27B abliterated, for the Ollama tag and usage warnings
• Huihui Qwen3.6-27B abliterated GGUF (lhca521), for the hf.co Ollama command
• Dolphin 3.0 Mistral 24B, for the system prompt and temperature guidance
• Apidog: best uncensored LLMs, 2026, for the llama.cpp version note on Qwen 3.8