Local model
Susi edits with one model, qwen3.5:4b-mlx, through Ollama on your Mac. There is no account, no API key and no server of ours: the text you rewrite goes to Ollama at 127.0.0.1 and nowhere else.
Set it up
- Install the Ollama app . The
ollamacommand-line tool alone works too, for example frombrew install ollama. - Download the model once:
ollama pull qwen3.5:4b-mlxThat is all. Susi talks to Ollama at http://127.0.0.1:11434 directly, bypassing any proxy settings.
What Susi does for you
- Starts Ollama. If Ollama is not running, the first edit starts it in the background and waits up to about twelve seconds: the Ollama app, or
ollama servewhen only the command-line tool is installed. A server Susi starts keeps running after you quit Susi, as the Ollama app does. - Says what is missing. Without Ollama, the edit stops at once, and Download Ollama on the line above the bottom bar opens the download page in your browser. Without the model, that line shows
ollama pull qwen3.5:4b-mlx: click it to copy the command, then run it in Terminal. ⌘K shows the same note above its field before you choose an action, and Turn into goal still works, since it needs no model. From the keyboard, Tab moves to the link or the command and Enter uses it. - Loads the model early. When you press ⌘K and Ollama is already running, Susi asks it to load the model into memory, so the edit itself starts sooner. This sends no text.
- Checks that the model is local. Before it sends any text, Susi asks Ollama about the model. If the model is missing, or Ollama would forward it to a cloud service, Susi stops with a message instead of sending your text.
Limits of one edit
- The text and the instruction together must fit within 12 KB. For a long prompt, select the passage to change.
- If the answer runs into the model’s context limit, Susi reports it and changes nothing.
- An edit gives up after about two and a half minutes. Esc cancels it any time before that.
- The model runs with temperature 0, so it does not add randomness of its own.
What the model keeps in your text →
Check Ollama yourself
curl --noproxy '*' -fsS http://127.0.0.1:11434/api/version # Ollama answers
ollama list # qwen3.5:4b-mlx is downloaded
ollama ps # what is loaded into memory nowSusi finds the Ollama app the way macOS does, and the ollama command in /opt/homebrew/bin or /usr/local/bin. A server Susi starts does not see the environment of your shell: if you set OLLAMA_* variables there, start the server yourself before you edit. Troubleshooting →