dwim

A coding agent that runs on your own GPU.

The agent, the model, and the engine that runs it, in one Rust program: no API key, no server, and nothing leaves your machine.

GitHub Model: Bonsai 2 27B

# install
$ cargo install --git https://github.com/dwim-sh/dwim

# start the shell in a project directory
$ dwim

# or ask one thing
$ dwim "what changed in the last commit?"

• A real coding agent. A shell where the model runs commands, reads their output, and keeps going until the task is done. It follows your project's AGENTS.md.
• A 27B model in 6 GB. Bonsai 2 27B, PrismML's ternary version of Qwen3.8-27B: every weight is -1, 0, or 1, so it fits on a laptop GPU.
• Its own inference engine. No llama.cpp, no CUDA. Metal on macOS, Vulkan everywhere else, and a CPU reference the GPU kernels are tested against.