Configure Reasoning Effort¶
Use the reasoning control in the composer to adjust how much thinking budget a model spends on its next reply. Available levels depend on the active model and are saved per chat.
Using the Reasoning Effort Slider in the Web UI¶
- Open a chat and open the model picker in the composer.
- Select a model that supports configurable reasoning.
- Move the Reasoning effort slider (or use the left/right arrow keys when focused) to pick a level.
- Send your message.
The control adapts to the selected model's catalog entry:
effortmode: Offers the model's supported effort steps (such aslow,medium,high,xhigh,max), plus Off when reasoning is optional (reasoning_mandatory: false).enabledmode: Offers an On / Off toggle when reasoning is optional, or stays enabled when mandatory.nativeandnonemodes: Use fixed provider behavior without an effort slider.
Lower effort levels reduce time-to-first-token for quick questions; higher levels give the model more room to plan complex code changes, math, or multi-step tool use.
Mandatory vs Optional Reasoning¶
Whether reasoning can be turned Off is controlled by reasoning_mandatory
in the model catalog. When you switch a chat to a different model, Numi
automatically normalizes the chat's saved reasoning setting to the closest level
supported by the new model.
Administrators can view or edit reasoning_efforts, reasoning_default_effort, and reasoning_default_enabled on the admin Models page (and use Refresh provider metadata for OpenRouter models). See Manage Models.
Programmatic Configuration via API¶
Set Reasoning Level¶
Update the reasoning level for a chat you own:
curl -X POST http://localhost:4567/api/v1/chats/YOUR_CHAT_UUID/reasoning \
-H "Authorization: Bearer YOUR_TOKEN" \
-H "Content-Type: application/json" \
-d '{"reasoning_level": "high"}'
Response:
message returns the normalized level that was saved (for example, requesting
"off" on a mandatory-reasoning model resolves to its default enabled effort).
Supported Effort Values¶
The chat endpoint accepts:
Always pass "off" (not "none") to disable optional reasoning via the Numi
API; the wire adapter translates "off" to the upstream provider's format (such
as reasoning: {"effort": "none", "exclude": true} on OpenRouter)
automatically.
Reasoning Modes¶
Catalog reasoning_mode |
UI & Adapter Behavior |
|---|---|
effort |
Exposes the model's reasoning_efforts list (plus off when reasoning_mandatory is false) |
enabled |
Exposes an on/off toggle when optional, or keeps reasoning enabled when mandatory |
native |
Model reasons natively without an effort parameter (for example, Z.AI or local <think> models) |
none |
Non-reasoning model |
Preserved Reasoning Across Turns¶
Reasoning effort controls how much the model thinks on the current turn, while preserve_reasoning controls whether returned reasoning traces are replayed in history on subsequent turns. See Preserved Reasoning and Model Catalog Reference for details.