Building AI features using Foundation Models. Streaming.

08 Oct 2025

We’ve already discussed the fundamental concepts of Foundation Models and their structured outputs. This week, we’ll delve into the process of streaming the partial results generated by the model.

Enhancing the Xcode Simulators.
Compare designs, show rulers, add a grid, quick actions for recent builds. Create recordings with touches & audio, trim and export them into MP4 or GIF and share them anywhere using drag & drop. Add bezels to screenshots and videos. Try now

We are already familiar with requesting and waiting for the results from Foundation Models. However, there are examples where we require real-time experience, where the intermediate output is displayed while the model is processing.

mport FoundationModels

@Generable struct Article {
    @Guide(description: "The title of the article")
    let title: String

    @Guide(description: "The content of the article")
    let body: String

    @Guide(description: "Useful tips related to the article")
    let tips: [String]
}

import Playgrounds

#Playground {
    let articleGenerationInstructions = "Write a health related article."
    let session = LanguageModelSession(instructions: articleGenerationInstructions)
    let response = session.respond(to: "Heart Rate", generating: Article.self)

    print(response.content)
}

As demonstrated in the example above, we define the Article type, which is annotated with the Generable macro. This macro enables us to receive the response from the Foundation Model in a specific format defined by the type. So, the Foundation Model generates an instance of the Article type and populates its properties with the guided content.

Here, we utilize the await keyword to wait for the final output. However, what if we desire to display the partial results and append additional content as soon as the Foundation Model generates it? For this specific scenario, Apple introduced the Streaming API for Foundation Models.

import Playgrounds

#Playground {
    let articleGenerationInstructions = "Write a health related article."
    let session = LanguageModelSession(instructions: articleGenerationInstructions)
    let stream = session.streamResponse(to: "heart rate", generating: Article.self)

    for try await article in stream {
        print(article)
    }
}

As you can see, all we need to do is use the streamResponse function instead of the respond function. The streamResponse function returns us an instance of the ResponseStream conforming to the AsyncSequence protocol and emitting the partial results of the Article type.

We define the Article type as a plain structure with properties using the let constant. All the magic here is hidden behind the Generable annotation. It automatically defines the PartiallyGenerated type inside the Article type, which defines the same properties, making them optional.

The order you define properties in the Article type plays a huge role. The Foundation Model respects the order you define properties and generates the title first, then the body and tips at the end. So, you should display them in the order you define them to make the user experience nice and appealing.

import Playgrounds

#Playground {
    let articleGenerationInstructions = "Write a health related article."
    let session = LanguageModelSession(instructions: articleGenerationInstructions)
    let stream = session.streamResponse(to: "heart rate", generating: Article.self)

    for try await article in stream {
        print(article.content)
    }

    let article = try await stream.collect()
    print(article.content)
}

You can always await the final result using the collect function on the ResponseStream type. It might be useful to collect the final results after the for loop.

Streaming transforms how we interact with Foundation Models, shifting from static responses to dynamic, real-time experiences. By leveraging the streamResponse API, we can progressively display model output as it’s generated — enhancing responsiveness and user engagement. I hope you enjoy the post. Feel free to follow me on Twitter and ask your questions related to this post. Thanks for reading, and see you next week!