Thinking / Reasoning Configuration

When adding or editing a self-managed model, admins can enable and configure thinking/reasoning behavior. This controls how the model reasons internally before producing an answer, and whether that reasoning output is fed back into subsequent conversation turns.

Why This Matters

Provider defaults for reasoning behavior can change between model versions. For example, a provider may ship a new model version with reasoning effort set to "none" by default when the previous version defaulted to "medium." Without explicit configuration, model quality can appear to degrade unpredictably. The thinking configuration gives admins explicit control.

Enabling Thinking

1. In the Add Model (or Edit Model) wizard, navigate to the Model Details step.

2. Toggle Enable Thinking/Reasoning to ON. When disabled (default), no thinking parameters are sent to the provider, and the remaining thinking controls are hidden.

3. When enabled, two additional controls appear: Thinking Budget and Propagate thinking in conversation.

Thinking Budget

The Thinking Budget control adapts to the selected provider. Available options, ranges, and defaults are driven by per-provider metadata so the wizard stays correct as providers add new models or change limits.

Provider	Control type	Options / Range	Default
OpenAI / Azure OpenAI / Azure AI Foundry	Dropdown	Low, Medium, High	Sourced from model metadata
Anthropic (direct API)	Numeric input	Token budget (min/max from model metadata)	From model metadata
Amazon Bedrock (Anthropic models)	Numeric input	Token budget (min/max from model metadata)	From model metadata
GCP Vertex AI (Gemini)	Numeric input + Dynamic option	Token budget within model range, or Dynamic (model decides its own budget)	Dynamic
OpenAI Compatible	Admin's choice	Effort dropdown (Low/Medium/High), numeric token budget, or Passthrough (lets the upstream service apply its own default)	Passthrough

Note on Anthropic models: For newer Anthropic models (Opus 4.6, Sonnet 4.6 and later), the thinking.type: "enabled" / budget_tokens pattern is deprecated. These models use thinking.type: "adaptive" with an effort parameter instead. Older models (Sonnet 4.5, Opus 4.5, etc.) do not support adaptive thinking and still require thinking.type: "enabled" with budget_tokens. Tabnine handles this distinction automatically based on the model version. Admins do not need to select between the two modes manually.

Propagate Thinking in Conversation

When thinking is enabled, a second toggle appears: Propagate thinking in conversation.

• ON (default): The model's thinking/reasoning output from previous turns is included as context in subsequent turns. This can improve multi-turn coherence for models that benefit from seeing their own reasoning chain.

• OFF: Thinking output is discarded between turns. Use this for models that degrade in quality when they receive their own reasoning output as input.

A tooltip on this toggle explains: "Some models may reduce quality when they receive their own reasoning output. Disable this if you notice degraded responses in multi-turn conversations with this model."