Clivern/Diffsay

Diffsay

Generate a Conventional commit message from the staged git diff, locally with MLX.

Requires Python 3.10+ and Apple silicon.

uv tool install diffsay==v0.2.0

Or from this repo:

uv tool install .

git add .
diffsay

In a terminal, the suggested subject is prefilled. Move with the arrow keys to edit it, then press Enter to git commit. Ctrl-C cancels.

Piped use, or --print-only, only prints the subject:

git commit -m "$(diffsay --print-only)"

Default model is Qwen2.5-Coder-7B-Instruct-4bit. Bare names resolve under mlx-community/.

diffsay -m Qwen2.5-Coder-14B-Instruct-4bit

Use a 7B+ instruct model. Sub-3B models (Gemma 1B, Qwen 1.5B) tend to invent tokens and ignore the “one line” rule.

Model	RAM (4-bit)	Notes
`Qwen2.5-Coder-7B-Instruct-4bit`	~6–8 GB	Default
`Qwen2.5-Coder-14B-Instruct-4bit`	~10–12 GB	Better whole-diff summaries
`Qwen2.5-Coder-32B-Instruct-4bit`	~20+ GB	Needs 24+ GB unified memory
`Llama-3.1-8B-Instruct-4bit`	~6–8 GB	US Model, follows instructions well
`Qwen3-8B-Instruct-4bit`	~6–8 GB	Newer general model
`Qwen3-14B-Instruct-4bit`	~10–12 GB	Stronger feat vs fix

uv sync
uv run ruff check src tests
uv run pytest