A reasoning system for Mistral 7b Instruct v0.3 (Or literally any LLM with a large enough context window and some simple modifications to the script)
Small desktop app (Tkinter) that runs a single Mistral-7B model locally and adds a multi-pass reasoning loop:
- Plan – model generates numbered high-level steps
- Execute – creates a hidden paragraph for each step
- Synthesize – merges internal thoughts into a draft answer
- Critique – self-reviewer lists flaws, or “OK”
- Rewrite – fixes draft if needed
The progress bar reflects each internal pass so you know the model hasn’t frozen. Prompts and responses are labelled in the chat pane.
You must have Mistral 7B Instruct v0.3 already saved to a folder on your computer. If you want to use a different LLM, simply modify the model loading area, the llm method that actually calls the LLM, and any formatting methods to make it match the required prompt formatting for whichever LLM you want to use.
You can even modify the llm method to send API requests to an online model instead of running one yourself!
git clone https://github.com/garagesteve1155/mistral-reasoning-gui
cd mistral-reasoning-gui
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python chat_reasoning_gui.py