Configuration options when accessing the Gemini API through Apple's Foundation Models framework

In each request to a model, you can send along some configuration settings to control how the model generates a response. Each Gemini model offers different configuration options.

Depending on the setting, the configuration is maintained for the lifetime of the geminiLanguageModel or the LanguageModelSession. If you want to use a different configuration, create a new model or session with that config.

The page describes the following configuration options:

• System instructions
• Thinking (aka "reasoning")
• Generated image output
• General configuration

Specify system instructions

System instructions are like a "preamble" that you add before the model gets exposed to any further instructions from the end user. It lets you steer the behavior of the model based on your specific needs and use cases.

Specify system instructions on the LanguageModelSession:

// ...

// Make sure you initialize your chosen Gemini API backend service
let model = FirebaseAI.firebaseAI().geminiLanguageModel(name: "GEMINI_MODEL_NAME")

// Specify the system instructions as part of creating the session
let session = LanguageModelSession(
    model: model,
    instructions: "You are a cat. Your name is Neko."
)

// ...

Configure thinking (aka "reasoning")

You can configure the amount of thinking (also called reasoning) that a Gemini model can do to generate its response. For details, best practices, and use cases for thinking, see the general Thinking guide.

Instead of thinking level, Apple uses reasoning level. You set the reasoningLevel on the LanguageModelSession:

// ...

// Make sure you initialize your chosen Gemini API backend service
let model = FirebaseAI.firebaseAI().geminiLanguageModel(name: "GEMINI_MODEL_NAME")

// Specify the reasoning level as part of creating the session
// Possible reasoningLevels: .light, .moderate, .deep, .custom("minimal")
let session = LanguageModelSession(
  model: model,
  contextOptions: ContextOptions(reasoningLevel: .REASONING_LEVEL))

// ...

Supported thinking level values

Note the following:

• Gemini 3.x models always use thinking; you cannot disable or turn off thinking for these models.
• Gemini 3.x models always use dynamic thinking – the model decides when and how much it thinks up to the configured amount.

The following table lists the thinking levels (reasoningLevel) that you can set for each Gemini model. Note that if you use a Gemini value (MINIMAL, LOW, MEDIUM, HIGH), it will be automatically mapped to the Apple version of that value.

Configure image output

General configuration

Give feedback about accessing the Gemini API through Apple's Foundation Models framework