Apple has released a native framework that lets developers build AI-powered apps using on-device language models without downloading weights, running servers, or paying per query. The Foundation Models framework, introduced at WWDC26 in June 2026, provides a unified API for accessing Apple's third-generation foundation models (AFM 3) alongside third-party providers, giving developers flexibility while maintaining privacy.

How Does Apple's Foundation Models Framework Work?

The framework operates through a simple three-step workflow: developers create a language model session, optionally describe the Swift type they want back, add tools the model may call, and stream the answer. The key innovation is that inference runs on the device for Apple's system models, eliminating download overhead and token-based billing entirely.

A developer can check availability, create a session with custom instructions, and ask questions in just a few lines of code. The framework includes "guided generation," a feature that turns free-form text into typed Swift values, making it practical for parsing receipts, forms, and screenshots without manual post-processing.

What Model Options Are Available to Developers?

The framework supports multiple backends, giving developers control over where inference happens and which model powers their app:

• AFM 3 Core: A 3-billion-parameter on-device model that serves as the default for the Foundation Models framework, optimized for speed and privacy on Apple Intelligence devices.
• AFM 3 Core Advanced: A 20-billion-parameter sparse model that activates only 1 to 4 billion parameters per request, natively multimodal, with weights stored in flash memory and loaded on demand.
• Private Cloud Compute models: Apple's server-based AFM 3 Cloud, ADM 3 Cloud, and AFM 3 Cloud Pro for tasks requiring more speed, image generation, and complex reasoning with tool use.
• Third-party providers: Since WWDC26, developers can plug in open-source models via MLXLanguageModel and CoreAILanguageModel, or cloud providers like Claude and Gemini through a pluggable LanguageModel protocol.

This flexibility means developers can start with Apple's on-device model for privacy-critical tasks, then switch to a server model or third-party provider for more complex reasoning without losing conversation context, using a feature called Dynamic Profiles.

What New Capabilities Did WWDC26 Add?

The June 2026 update expanded the framework significantly. Developers can now pass images alongside text, allowing the model to answer questions about photos. Vision framework tools such as optical character recognition (OCR) and barcode reading can be exposed to the model as callable tools. Dynamic Profiles let apps switch the model, tools, and instructions within a continuing session, enabling seamless transitions from quick on-device answers to server models without losing context.

Apple also released a Python SDK and a command-line tool called "fm" for scripting prompts against the system model on a Mac, plus an Evaluations framework for testing prompts before release. This allows developers to prototype prompt designs quickly and port winning prompts into Swift production code.

How Does AFM 3 Compare to Other Apple AI Models?

Apple distinguishes between Foundation Models and its open-source research models. FastVLM, MobileCLIP2, Depth Pro, AIMv2, and SHARP are downloadable models that run anywhere: in browsers, Python environments, MLX, or on older devices. Foundation Models, by contrast, ship with the operating system and are accessed through the framework on Apple Intelligence devices, offering zero download overhead and system-level privacy guarantees.

The AFM 3 family represents Apple's third generation of foundation models. All models are multimodal, meaning they can process both text and images. Apple reports gains in text quality, image understanding, and speech synthesis compared to earlier versions. Training data includes public, licensed, and synthetic sources; Apple states that no private user data is used.

Core ML and its WWDC26 successor, Core AI, operate differently. Those frameworks run models that developers convert and ship themselves. The Foundation Models framework is a higher-level session API over models Apple provides, or since WWDC26, over any provider that conforms to the LanguageModel protocol.

What Are the Practical Implications for App Developers?

The framework eliminates several traditional barriers to shipping AI features. Developers no longer need to download and bundle model weights, reducing app size and installation friction. On-device inference means no server infrastructure to maintain, no token-based billing, and no latency from network round trips. For users, this translates to faster responses and stronger privacy, since data never leaves the device for Apple's system models.

The pluggable architecture also reduces vendor lock-in. A developer can prototype with Apple's on-device model, then swap in Claude, Gemini, or an open-source model if requirements change, all through the same API. This flexibility encourages experimentation and lets teams choose the best tool for each task without rewriting code.

Apple announced that the Foundation Models framework, including Linux support, would be open-sourced in summer 2026, further expanding its reach beyond Apple platforms.

When Should Developers Use Foundation Models vs. Other Approaches?

The framework is best suited for apps running on Apple Intelligence devices where developers want zero-download, system-managed inference. If an app needs a specific open model, cross-platform reach including the web and Linux, control over quantization and latency, or support for older devices, FastVLM or similar downloadable models are better choices.

For developers building on Apple platforms, the Foundation Models framework removes friction around model distribution and cost. The combination of on-device privacy, zero download overhead, and support for multiple backends makes it a compelling option for shipping AI features quickly without infrastructure overhead.