Model and effort per chat
In Perch Desktop each chat keeps its own model and its own effort level, so a cheap chat and a deep one can sit side by side without you resetting anything.
Picking a model used to be one setting for the whole app: choose a model in one chat and every chat changed with it. That is no longer how it works. In Perch Desktop the choice belongs to the chat, and so does the effort level.
The short version
How a chat gets its model
Three ways, in order:
- You pick one in that chat, from the model button in its composer or with
/model. - Its first turn. A chat with no model of its own keeps the one its first turn ran on, so a pick you make somewhere else afterwards never moves it.
- Its saved row. The chat's model is stored with the chat, so it is still there when you come back to it days later.
A chat you have not written in yet follows the app's choice until its first turn. A chat written in before per-chat models existed has no saved model, and runs on the app's choice as it stood when you opened the window, rather than following picks made in other chats afterwards.
What you can pick depends on your plan. Starter can pin the models in its own pool; Pro can pin any model except the Ultra-only ones; Ultra can pin the full registry. See Models.
Effort is separate
The effort level is tracked apart from the model. Changing a chat's model never rewrites its effort, and a chat remembers both independently, by the same three routes: a pick on its effort control, its first turn, or its saved row.
The control itself is a bar in the model menu, and its stops are the levels the selected model supports.
| Setting | What it means |
|---|---|
| Off | No thinking. Fastest, and lightest on your usage. |
| Low | Shallow reasoning. Short scoped tasks where speed matters. |
| Medium | Balanced. Fewer, more consolidated tool calls. |
| High | Default. Complex reasoning and agentic work. |
| XHigh | Deeper reasoning for difficult, multi-step work. |
| Max | Maximum supported reasoning for quality-first tasks. |
| Ultra | Deepest mode, available only when the selected model advertises it. |
Choosing a level is part of Pro and Ultra, and a Starter account holding an Unlimited DeepSeek pass can choose one while the pass runs. Otherwise Starter turns run at High.
Models express reasoning in completely different ways: some ignore reasoning parameters entirely, some take a single on and off switch, and others publish their own named ladder. The level you pick is coerced onto whatever the resolved model actually understands, and an unsupported level lands on the nearest one below it rather than above. The harness covers that in more detail.
Helpers follow the chat
The model for the helpers Perch spawns during a turn defaults to Inherit, which follows the chat: a model you pinned for a chat is used for its helpers too, and on Roost they stay on automatic routing. Naming a different model is how you get one model checking another's work. See Sub-agents.
In the CLI
The CLI keeps one model and effort per session rather than per chat. /model shows or sets it, /effort shows or sets the level, and /status prints both along with the mode, the persona and the working directory.
The Chats tool
On Pro and Ultra, the agent can see your other chats, read one, send it a message or stop its turn, so one piece of work can hand off to another without you ferrying messages between them.
Local models
Run Perch on a model on your own computer through Ollama, LM Studio or a llama.cpp server. Turns on it are not metered, and the address never leaves your machine.

