Localight for iOS

Localight is a simple SwiftUI chatbot app for iOS 26, powered entirely by Apple’s on-device Foundation Models. Designed for demonstration purposes, Localight offers fast, private, and completely offline AI chat — no internet connection or server required.

Localight showcases how to integrate Apple's local LLM into a native iOS experience using SwiftUI and the new Foundation Models framework.

⚠️ Disclaimer: This app is for demonstration purposes only and is not production-ready. Output generated by the local large language model may be inaccurate, incomplete, or misleading.

✨ Features

• 🧠 On-device LLM: Uses Apple’s local Foundation Models for text generation.
• 🔐 Privacy-first: All conversations stay on your device. No data is sent to the cloud.
• ⚡ Fast & Offline: No internet needed. Responses are generated locally.
• 💬 Minimalist Chat UI: Clean SwiftUI interface for interacting with the model.
• 🗑️ No history: Conversation is not saved after closing the app.

🛠 Manual

• Import the Library: To work with Foundation Models, you must import the library in every file where you intend to use them. Go with:

  import FoundationModels

• Check Availability: The key object is the SystemLanguageModel. This can indicate whether the model is .available or .unavailable. In case of unavailability, a reason is also provided.

• Create a LanguageModelSession: To start prompting, you need to create a LanguageModelSession. When you create a session you can provide instructions that tells the model what its role is and provides guidance on how to respond. These instructions should never be editable by the user:

  let session = LanguageModelSession(instructions: "You are the best friend.")

• Generate a Response: Finally, call the method to receive a result as a String by using:

  let response = try await session.respond(to: promptAsString).content

• Stream a Response: Otherwise, you can call the method to receive a result as a String that is streamed by using:

  let stream = session.streamResponse(to: promptAsString)
  
  do {
      for try await chunk in stream {
          self.streamingResponse = chunk.content
      }
      
      let response = try await stream.collect().content
  } catch {}

📏 Context Window & Token Limits

Apple’s on-device Foundation Models operate with a limited context window per session. The context window defines how many tokens the model can process within a single LanguageModelSession.

• A token is a unit of text processed by the model.
• In Western languages (e.g. English or German), 1 token ≈ 3–4 characters.
• In East Asian languages (e.g. Japanese or Chinese), 1 token ≈ 1 character.
• The system model currently supports up to 4,096 tokens per session.

If this limit is exceeded, the framework throws the following error: LanguageModelSession.GenerationError.exceededContextWindowSize(_:)

For more details, see Apple’s official documentation: TN3193 – Managing the on-device foundation model’s context window