What's new in the Foundation Models framework?

The Foundation Models framework got a lot of new capabilities this year, including:

• Open source framework and utilities

  • The core framework is going open source.
  • A new foundation_models_framework_utils package adds emerging and experimental building blocks between OS releases.

• New on-device model

  • A rebuilt on-device model that’s better at intelligence, logic, and tool calling.
  • New APIs for inspecting context size and token counts in prompts, instructions, and transcripts.

• Vision support

  • Support for images in prompts and tool calling with images.

• Private Cloud Compute and broader model support

  • Access to server models and a more flexible model abstraction layer.
  • The framework is expanding to support more model providers and local/server model options.

• System tools

  • New tools powered by Spotlight and Vision to enhance sessions.

• Dynamic Profiles

  • A new primitive for building agentic experiences with flexible, composable behavior.

• Evaluations framework

  • New Swift framework for measuring prompt quality and validating AI features.

• Command-line and scripting tools

  • The new fm command-line tool for using models from Terminal.
  • A Python SDK for using the same model capabilities in Python.

• Open-source ecosystem

  • Utilities for transcript management, skills, and server-backed model access.
  • The framework is being positioned to work across Swift environments, including server-side use.

If you want, I can also summarize the changes by chapter, such as models, tools, agentic apps, or open source.