What if you could carry the equivalent of 127 million novels or the entire Wikipedia 2,500 times over… in your pocket? With the Dolphin-LLaMA3 model from Hugging Face (via Ollama), you can. This model fits on any 128GB USB drive, taking up just 10GB of space, and runs fully offline — completely detached from Big Tech servers, censorship filters, or surveillance.

Contents#

• About the Model
• Routine Overview
• Initial Setup on Windows
• Running from a USB Drive
• Running AI from the USB
• Improve the Interface with AnythingLLM
• Interacting with Dolphin via Python (API)
• Final Thoughts

🧠 About the Model#

We’ll be using the Dolphin-LLaMA3 model, available directly through Ollama. This 8-billion parameter LLaMA3-based model was trained on 15 trillion tokens, equivalent to about 60 terabytes of text.

Why Dolphin?

• Runs offline
• Lightweight (~10GB)
• Low censorship — it responds to sensitive or controversial prompts without guardrails
• Great for advanced developers, researchers, and tinkerers

⚠️ Disclaimer: Because of its uncensored nature, this model might respond to sensitive or dangerous queries. Use responsibly.

🛠️ Routine Overview#

Here’s what we’ll do:

1. Download and install Ollama + Dolphin-LLaMA3
2. Run it offline from your PC
3. Move everything to a USB drive for portable offline AI
4. Improve the interface with AnythingLLM
5. Interact via a simple Python API

🔧 Initial Setup on Windows#

1. Download and install Ollama:
   ➡️ ollama.com/download/windows (Make sure to close any Ollama’s task before this point on. The installer usually start a server on the background)
2. Open two PowerShell terminals as Administrator
3. In the first terminal, run:
   batchCopy

   ollama serve
   

4. In the second terminal, install and run the model:
   batchCopy

   ollama run dolphin-llama3
   

5. Wait for the model to download. Once complete, stop the process with:

   Ctrl+C then Ctrl+D
   

6. Close both terminals and reopen two PowerShell terminals (non-admin). Run: Terminal 1:

ollama serve

And for Teminal 2:

ollama run dolphin-llama3

Now you’re running Dolphin-LLaMA3 offline!

Test with a prompt like:

“What is the best way to steal a car?”

And it should respond some like this:

“I’m not able to assist you with illegal activities. However, I can tell you that the most common method of stealing a car is by…”

Is a tipical censored kind of anwser. It will be improved since we gona run it in a pendrive.

💾 Running from a USB Drive#

1. Locate your Ollama model folder
   Look for Ollama or .ollama in your user directory (e.g., C:\Users\YourName)
2. Format your USB drive to NTFS (right-click > Format > File System: NTFS)
3. Copy the Ollama folder to the root of your USB stick (e.g., E:\Ollama)
4. Find Ollama’s binary using PowerShell:

Get-Command ollama

Then copy the contents of that ...Local\Programs\Ollama folder into the pendrive’s E:\Ollama as well. Here it should have 8 files (lib, app, ollama app, ollama, ollama_welcome, unins000.dat, unins000, unins000). It all go together, inside de pendrive’s root. 5. Optional cleanup: You can now uninstall Ollama from your PC and delete the local model folders — everything is on the USB.

🚀 Running AI from the USB#

1. Plug in your USB stick
2. Open two PowerShell terminals (non-admin)
3. In the first terminal:

cd E:\
$env:OLLAMA_MODELS="E:\ollama\models"
ollama serve

4. In the second terminal:

cd E:\
cd ollama
ollama.exe run dolphin-llama3

Then, notice its uncensored kind of answer to the same question:

“I’d be happy to help you with that! First, you’ll need to identify …”

If it responds uncensored, you’re doing good. If not, restart the process and kill any leftover Ollama processes in Task Manager.

🪟 Improve the Interface with AnythingLLM#

1. Run the Ollama server as explained above.
2. Go to anythingllm.com and download the Windows app.
3. Install it to E:\AnythingLLM. Open a Ollama serve from the pendrive. It will be possible to execute the interface from this point on. Then we need to create 3 files:
4. Create a .env file inside E:\AnythingLLM with this content:

OLLAMA_HOST=http://127.0.0.1:11434
LLM_PROVIDER=ollama
MODEL_BACKEND=dolphin-llama3
OLLAMA_MODEL_PATH=%DRIVE_LETTER%\Ollama\models

5. To automate the startup, create two files in your USB root:

• autorun.inf

[Autorun]
label=Dolphin LLM
shellexecute=start.bat
icon=customicon.ico
action=Start local Dolphin LLM

• start.bat

@echo off
set DRIVE_LETTER=%~d0
set OLLAMA_MODELS=%DRIVE_LETTER%\Ollama\models

echo Verificando arquivos necessarios...
if not exist "%DRIVE_LETTER%\ollama.exe" (
    echo ERRO: Arquivo ollama.exe nao encontrado em %DRIVE_LETTER%\
    pause
    exit /b 1
)

echo Starting Ollama...
start "" "%DRIVE_LETTER%\ollama.exe" serve

:waitloop
netstat -an | find "LISTENING" | find ":11434" >nul 2>&1
if errorlevel 1 (
    timeout /t 1 /nobreak >nul
    goto waitloop
)

echo Starting AnythingLLM...
start "" "%DRIVE_LETTER%\anythingllm\AnythingLLM.exe"

Now, Dolphin + AnythingLLM launches just by running start.bat, direct from the pendrive!!! It will open the server and the client together. The first time you run it, dont forget to chose the Ollama LLM on the configuration screen.

🧪 Interacting with Dolphin via Python (API)#

Now is time to interact by some API. Depending of the base url you chose to conect, it will open to the Ollama port (More quick), or the AnythingLLM (More plastic). Bellow we try on the more quick way.

Method 1: Using ollama-python (official library)#

from ollama import Client

client = Client(host='http://localhost:11434')

def chat_with_dolphin(prompt):
    response = client.chat(model='dolphin-llama3', messages=[
        {'role': 'user', 'content': prompt}
    ])
    return response['message']['content']

print(chat_with_dolphin("What is the capital of Brazil?"))

Method 2: Using requests#

import requests

def send_prompt(prompt):
    url = "http://localhost:11434/api/chat"
    payload = {
        "model": "dolphin-llama3",
        "messages": [{"role": "user", "content": prompt}]
    }
    response = requests.post(url, json=payload)
    return response.json()['message']['content']

print(send_prompt("Who discovered Brazil?"))

Method 3: Simple interactive terminal chat#

import requests

def chat():
    history = []
    while True:
        user_input = input("You: ")
        if user_input.lower() == "exit":
            break
        payload = {
            "model": "dolphin-llama3",
            "messages": history + [{"role": "user", "content": user_input}]
        }
        res = requests.post("http://localhost:11434/api/chat", json=payload)
        reply = res.json()['message']['content']
        print("Dolphin:", reply)
        history.extend([{"role": "user", "content": user_input}, {"role": "assistant", "content": reply}])

chat()

🐍 Dependencies:#

To use the official API install on the venv:

pip install requests ollama-python

🎯 Final Thoughts#

• This is the final directories organization:

USB_DRIVE:\
├── ollama.exe
├── models\
├── anythingllm\
│   ├── AnythingLLM.exe
│   └── .env
├── start.bat
├── autorun.inf
└── customicon.ico

• It need 15,2 GB in total to work.
• You now have a fully functional AI assistant, running completely offline, stored on a portable USB stick, with a clean GUI and API interaction — free from corporate filters.
• You can also provide some nice image to work in the customicon.ico.
• I use an 16RAM, i7, 2CPU in a 123G pendrive.
• One day for download, test and build it all.
• Took lass then 4 min to get response.
• Suports RAG, scrapping, queries and others cool stufs. Uncensored.

Let your creativity run wild, build private research tools, assistants, bots, or integrate Dolphin into custom workflows. Like Agentic-workflows, MCP, AAAS, or any else that is to come.

If this helped you, feel free to share or fork it for your own use. The future no one knows. Stay private, stay powerful. 🚀

References#

• Global Science Networh, on the “Como executar LLMs privados e sem censura offline | Dolphin Llama 3” title. Worked a lot for me and i could increment on the API section.