dwim is an AI coding agent that runs on your own GPU. Give it a task and it reads your code, edits files, and runs commands.
Open source. No account, no API key, no token bill.
macOS on Apple silicon · Linux with a Vulkan GPU · MIT
dwim is experimental software. Expect bugs, rough edges, and changes that break things from one commit to the next. The model runs shell commands on your machine without asking first, and dwim does not sandbox them, so run it in a container or a virtual machine, on code you have committed or backed up.
Run dwim to open an interactive shell and work through a task with the agent.
Or pass a prompt on the command line. dwim prints only the reply to standard output, so you can redirect it or call it from your scripts.
$ dwim
$ dwim "Explain the locking assumptions in this module."
$ dwim "Add a regression test for this failing case."
$ dwim "Find the allocation hot spots in this benchmark."Your code and prompts stay local. Work on a plane or on code you can't upload. Once the model is downloaded, dwim works offline.
Run it as often as you need without watching a token budget.
Bonsai 2 27B reads files, runs tools, and follows your project's AGENTS.md. Its 27 billion parameters fit in about 6 GiB of memory.
PrismML compressed it from Qwen3.8-27B, which needs 50 GiB of memory.
Yes, if your GPU has 8 GiB of memory. The model takes about 6 GiB, and the conversation needs the rest. An AMD card from 2019 is enough.
You can even run it on the CPU, slowly.
Install dwim, open a repository, and give it a task.