Foundation Models updates

Learn about important changes to Foundation Models.

Overview

Browse notable changes in Foundation Models.

Build multimodal agentic app experiences by using the LanguageModelSession.DynamicProfile API.
Use the improved error types, like LanguageModelError for model-specific errors, SystemLanguageModel.Error for on-device Apple Foundation model errors, and LanguageModelSession.Error for errors related to the session but not the model.

Use the latest on-device SystemLanguageModel that follows instructions more accurately and produces better results, including in complex scenarios. Because the model changes when a person updates to iOS 27, iPadOS 27, macOS 27, and visionOS 27, test your prompts with the new model to verify your app’s behavior.
Adopt the LanguageModel protocol to use any large language model — server or on-device — with the Foundation Models framework.
Use PrivateCloudComputeLanguageModel to access more reasoning capabilities and a larger context size.
Perform image analysis tasks by including an image in your prompt and using tools the Vision framework provides, like OCRTool and BarcodeReaderTool.

Control how the model interacts with tools for your request by using GenerationOptions.ToolCallingMode.

Use the updated Analyzing the runtime performance of your Foundation Models app to get detailed insight into the complex workflows you build. The instrument provides insight into latency, prompts sent to the model, model output, tools and token usage, and so on.

Get the Foundation Models framework utilities to access a collection of building blocks to help you explore emerging practices in working with large language models.
Use CoreAILanguageModel and MLXLanguageModel to integrate on-device models with the Foundation Models framework.

Use the Foundation Models SDK for Python to access the on-device foundation model at the core of Apple Intelligence.

Use the latest on-device large language model that improves instruction-following and tool-calling abilities. Because the model changes when a person updates to iOS 26.4, iPadOS 26.4, macOS 26.4, and visionOS 26.4, test your prompts with the new model to verify your app’s behavior. If necessary, update and maintain prompts for each model version.
Reduce the possibility of blocking benign content with improved guardrails for SystemLanguageModel.
Measure how many tokens your prompt, instructions, or entire session transcript uses with tokenCount(for:).
Use the contextSize property to get the maximum context size — in tokens — that the SystemLanguageModel supports.
Use the #Playground macro in Xcode to view an estimate of the usage of 4,096 tokens in the available context window. When you run the canvas, the output displays Input Token Count and Response Token Count separately.