DeepSeek Coder CLI

A lightweight terminal chat interface for DeepSeek-Coder-6.7B-Instruct using llama.cpp. Supports streaming responses, multi-turn conversations, file attachments, and CPU performance benchmarking.

Requirements

• Python 3.10+
• llama.cpp (built with llama-server and llama-bench)
• DeepSeek-Coder-6.7B-Instruct GGUF model
• rich (only for deepseek_chat.py; deepseek_chat_no_style.py needs no packages)

Usage

Start an interactive chat:

python3 deepseek_chat.py

Send a one-shot prompt:

python3 deepseek_chat.py -p "Explain binary search trees"

Pipe a file into the model:

cat app.py | python3 deepseek_chat.py -p "Review this code"

Run a performance benchmark:

python3 deepseek_chat.py --bench

Keep the server running:

python3 deepseek_chat.py --serve-only

Quick launch

ds-chat.sh sets up a virtual environment, installs rich if needed, and starts the chat (For this, please use deepseek_chat.py instead of deepseek_chat_no_style.py). Add an alias to your .zshrc or .bashrc:

alias ds-chat='~/llm/scripts/ds-chat.sh'

Interactive Commands

Command	Description
/file PATH	Attach a file to the next message
/system [TXT]	View or change the system prompt
/temp X	Change temperature
/max N	Set maximum response tokens
/reset	Clear conversation history
/save PATH	Save conversation as Markdown
/help	Display available commands
/exit	Exit the application

Configuration

Command-line options allow customization of context size, CPU threads, KV cache precision, sampling parameters, and server arguments. Run:

python3 deepseek_chat.py --help

Files

• deepseek_chat.py - styled version (Markdown rendering via rich)
• deepseek_chat_no_style.py - plain text version (standard library only)
• ds-chat.sh - venv + launch wrapper

The code is commented with Doxygen blocks.