Diffusion models in the terminal

Most language models write an answer one token at a time. A diffusion language model starts from a rough version of the whole answer and refines it in a few passes. CLIO Coder can use Inception's Mercury models this way, for chat and for workers, with tool calling.

Connect Mercury

Run guided setup:

clio-coder configure

On a first run, choose Guided setup. If Clio lists routes it found, pick Set up something else. Choose An API key and type $INCEPTION_API_KEY; setup reads the provider from the variable name, picks mercury-3, and sends one short test message before saving. When you add Mercury to an existing setup instead, choose A provider account or API, then Inception.

Or add the target directly:

clio-coder targets add --id mercury --runtime inception --model mercury-3 --api-key-env INCEPTION_API_KEY

The runtime lists mercury-3, its default, mercury-2.5, mercury-2, and mercury-edit-2. Setup checks the model you name against the provider's model list and refuses one that is not on it unless you pass --force. mercury-decide answers typed decisions for System One only and cannot be a chat model.

What you see

In the terminal, each update from the model carries the whole answer so far. CLIO Coder replaces the visible text on every frame instead of appending to it, and dims the part that is still changing. There is no progress gauge.

Tool calls, permission cards, /view, and run records work as they do with any chat model. Headless runs, workers, the desktop app, and editors connected over ACP receive an ordinary stream of text.

Where it fits

A quick whole-answer model suits short, bounded requests: a commit message, an explanation of one function, a first draft to compare. /draft uses the active chat model, so with Mercury active it returns several candidates quickly. You can also keep another model for conversation and route selected workers to the Mercury target.

Try it on a task you can check before trusting it with a long one. This page makes no speed or quality claim; measure both on your own repository.