ThinkStick CLI

Offline AI · On a Drive. (now also: offline AI, on a pip install)

This is the trimmed-down CLI edition of ThinkStick, built by DarkSyntax. The full ThinkStick product (USB drive / ThinkStick Pro) ships three specialist models (research, coding, math) behind a terminal-styled browser UI. This CLI is deliberately smaller: one model, one process, no server, no browser tab.

• No account, no API key, no cloud
• One llama.cpp model, downloaded once from Hugging Face, then never touched over the network again
• Runs entirely on your machine's own CPU (or GPU, if you enable offload)
• Nothing is installed system-wide beyond the Python package itself

THINKSTICK  cli edition
Plug in. Boot up. Think offline.

[ OK ] checking engine
[ OK ] checking model file
[ OK ] loading model .......... Qwen2.5-Coder-7B-Instruct-Q4_K_M.gguf
[ OK ] local inference ........ this machine only
[ OK ] all systems ready

you > is this actually offline?
thinkstick > yes. turn off your wifi and ask me again.

Install

pip install thinkstick

First run downloads the default model (~4.7 GB, one time, needs internet). Every run after that is 100% offline.

Usage

# start chatting (auto-downloads the model on first run)
thinkstick

# same thing, explicit
thinkstick chat

# just fetch the model, don't start a chat
thinkstick download

# see what model/config is currently active
thinkstick info

# use a different GGUF model from Hugging Face
thinkstick chat --model-repo bartowski/Qwen3.5-4B-Instruct-GGUF \
                 --model-file Qwen3.5-4B-Instruct-Q4_K_M.gguf

# push work onto a GPU (llama-cpp-python must be built with GPU support)
thinkstick chat --gpu-layers 20

# bigger context window, longer replies
thinkstick chat --ctx 8192 --max-tokens 2048

Type exit, quit, :q, or hit Ctrl+C to leave a chat session.

Why only one model

The full ThinkStick build (USB / ThinkStick Pro) is designed for MSME and clinic deployments where swapping between a research, coding, and math specialist matters. This CLI exists for a simpler use case: a single pip install that gets you a private, offline model with the least moving parts possible. It uses llama-cpp-python directly instead of a separate llama.cpp server process, so there's no port, no CORS, and no second terminal to manage.

Where things live

• Downloaded models: ~/.thinkstick/models/ (override with THINKSTICK_HOME)
• Default model: bartowski/Qwen2.5-Coder-7B-Instruct-GGUF, file Qwen2.5-Coder-7B-Instruct-Q4_K_M.gguf (override with THINKSTICK_MODEL_REPO / THINKSTICK_MODEL_FILE, or the --model-repo / --model-file flags)

License

MIT © 2026 SANJJAY — see LICENSE.

Built by DarkSyntax · darksyntax.xyz · sanjjay.stars@gmail.com