Model roles
Choose which model handles each kind of work, from your main conversations to memory, images, and speech.
Not every job needs your biggest model. Settings → Model Roles lets you assign a model to each kind of work, so routine background jobs can run on a small, cheap model while your conversations use the best one you have.
| Role | Used for | If left empty |
|---|---|---|
| Primary | Your conversations. Needs tool use and a large context window. | Required |
| Cheap | The default for lightweight background jobs. A small, fast model is ideal. | Primary |
| Title | Naming your conversations. | Cheap |
| Input Suggestion | Guessing your next message, shown in grey in the chat box. Open the row to turn the suggestions off. | Cheap |
| Memory Extraction | Picking out facts worth remembering after each turn. | Cheap |
| Decision | Quick yes-or-no calls: whether a goal is done, whether a tool call is safe to approve automatically, and, after each reply in a project conversation, whether it is waiting on you. | Cheap |
| Goal Judge, Permission Review | Override Decision for just one of those calls. Found under Advanced decision roles. | Decision |
| Memory organizer | Overnight memory consolidation. | Primary |
| Vision | Reading images. | Primary, if it can see images |
| Embedding | Memory search. | Nothing; assign one to use memory search |
| Image Generation, Image Editing, Video generation, TTS | The creative tools. | Nothing; assign one to use the tool |
| Decision model | Optional. A decision model that routes new kinds of Mailbox events. | Decision does it |
Decision models
A decision model doesn't write text: it picks from fixed choices and says how sure it is, in a few hundred milliseconds and for a fraction of a chat model's cost. Suzent can use:
- OpenAI:
openai/gpt-6-luna, through OpenAI's Decisions API. - Gemini:
gemini/gemini-decision-1, through the Gemini Decisions API. - A LiteLLM proxy you run: its decision models, such as Cloudflare's or
Databricks', as
litellm_proxy/<model>. - TypeSafe's Jev or a self-hosted decision model that speaks the System One protocol, added as a custom provider (see Model capabilities).
Each uses the key and address you set for that provider in Settings → Providers. Pick one under Decision model in Settings → Model Roles. Today it routes the first event of each new kind in the Mailbox. When it isn't sure of its choice, the event goes to Regent and no rule is saved, so the next event of that kind is decided again. If it fails or isn't set, the Decision role does the job instead.
A decision model can't chat, so it is never offered for the other roles.
Backups
Each role can also hold backup models. If the first one fails, Suzent tries the next in order.