Nautilus

Nautilus is a small terminal assistant, 0.8B parameters. You describe what you want in plain words, and it replies with the command to run. It works on Linux, macOS and Windows (PowerShell), and it's small enough to run on almost any machine: the q4_k_m file is about 530 MB.

Accuracy in %, full benchmarks run through their official harnesses, with every command executed and checked. The general models used each benchmark's standard prompt; Nautilus used its one-line system prompt.

What it does

Every reply is exactly one of three things:

Reply	When
a bare command, e.g. du -sh ~/Downloads/* | sort -h	the request can be done in the terminal
ANSWER: <one to three sentences>	the job is done, or the question needs an explanation rather than a command
UNCLEAR: <one question>	the request is ambiguous, or running a guess could do damage

When you send the command's output back to it, it continues from there: the next command, or an ANSWER: once the task is done.

What it doesn't do

• It isn't a chat model. It won't hold a conversation, write prose, or answer questions unrelated to the terminal.
• It proposes commands but doesn't run them. Always read a command before you run it, especially anything that deletes, overwrites or changes permissions.
• It gives one command per turn, not multi-line scripts.
• It can get rare flags wrong. It's reliable on common tools but can invent a flag for less common ones.

How to use it

Use one short line as the system prompt, describing the machine. Nautilus was trained on this format:

nautilus on <OS> (<shell>), <user>@<host>, home <home directory>.

For example:

nautilus on Ubuntu 24.04 LTS (bash), sam@lab, home /home/sam.
nautilus on macOS 15 Sequoia (zsh), sam@lab, home /Users/sam.
nautilus on Windows 11 Pro (PowerShell 7), sam@lab, home C:\Users\sam.

Then send the request as the user message, and send the terminal output back as the next user message. Use temperature 0 and turn thinking off.

llama.cpp

llama-server -m nautilus-q4_k_m.gguf --jinja --reasoning off -c 8192

curl http://localhost:8080/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "messages": [
    {"role": "system", "content": "nautilus on Ubuntu 24.04 LTS (bash), sam@lab, home /home/sam."},
    {"role": "user", "content": "which folders in my home take the most space?"}
  ],
  "temperature": 0
}'

A typical exchange:

user:      I want to move the pngs in ~/Pictures/2026 into a folder called shots
nautilus:  ls ~/Pictures/2026; test -d ~/Pictures/shots && echo exists || echo missing
user:      IMG_0431.png  IMG_0432.png  notes.txt
           missing
nautilus:  mkdir -p ~/Pictures/shots && mv ~/Pictures/2026/*.png ~/Pictures/shots/
user:      sam@lab:~$
nautilus:  ANSWER: Moved the two pngs into ~/Pictures/shots.

Files

This repo holds the safetensors weights. The GGUF builds (q4_k_m, q8_0, f16) are in Loke-60000/nautilus-gguf.