major
Evaluations framework for AI feature quality measurement
A new Swift framework (available in Xcode 27, supported on iOS, macOS, watchOS, visionOS) provides structured evaluation pipelines with Evaluation protocol, Metric, Evaluator, ModelJudgeEvaluator, ScoreDimension, TrajectoryExpectation, ToolCallEvaluator, SampleGenerator (with synthetic data generation), and MetricsAggregator.custom() — formalizing a hill-climbing develop/run/analyze loop for AI feature development.
major
Core AI framework for custom on-device model deployment
Core AI is a new memory-safe Swift framework for deploying any custom model — from compact vision models to multi-billion parameter LLMs — on Apple Silicon, featuring Python-based PyTorch conversion (coreai-torch), config-driven compression (coreai-opt), ahead-of-time compilation (xcrun coreai-build / AIModelCache), stateful KV-cache support, custom Metal 4 kernel embedding via TorchMetalKernel, a visual debugger, multi-function .aimodel assets, and zero server dependencies.
major
Foundation Models framework — multimodal, server model support, open source, and Linux
The Foundation Models framework gains multimodal image input (UIImage, NSImage, CGImage, Core Image, pixel buffers), a unified LanguageModelSession that works with on-device, Private Cloud Compute, MLX, Core AI, and third-party models via the new LanguageModel protocol, and is being open-sourced with a Swift server API that also runs on Linux. A Python SDK (apple_fm_sdk) and an open-source Swift utilities package with emerging agentic patterns are also shipping.
major
PrivateCloudComputeLanguageModel API with reasoning levels and 32K context
A new PrivateCloudComputeLanguageModel type gives developers no-account, no-API-key access to Apple's larger Private Cloud Compute model (32,000-token context window, including on watchOS 27), with configurable reasoning levels (.light and .deep via ContextOptions(reasoningLevel:)) and token usage reporting including reasoningTokenCount.
major
Dynamic Profiles API for agentic model orchestration
A new DynamicProfile protocol provides a declarative, SwiftUI-style way to manage per-prompt model, instructions, tools, and context for a LanguageModelSession, enabling multi-mode agentic apps that swap models mid-session while preserving conversation history; baton-pass and phone-a-friend multi-model orchestration patterns are formally named and supported.
major
MLX distributed training and inference across multiple Macs via RDMA over Thunderbolt
Starting in macOS 26.2, RDMA over Thunderbolt 5 enables high-bandwidth direct memory transfers between Macs; MLX and MLX LM leverage this (plus the open-source JACCL collective communication library) for tensor-parallel and pipeline-parallel LLM inference and data-parallel fine-tuning across multi-Mac clusters, achieving near-3x throughput on 4-Mac setups and enabling trillion-parameter models. mlx.launch and mlx.distributed_config CLI tools orchestrate the cluster.
major
Image Playground — photorealistic generation on Private Cloud Compute
Image Playground's generative model is rebuilt on Private Cloud Compute and now produces high-quality photorealistic images in virtually any style from photos or natural language descriptions, with data never stored or shared even with Apple.
major
Private Cloud Compute access for App Store Small Business Program members
Developers in Apple's App Store Small Business Program with fewer than 2 million downloads can access next-generation Apple Foundation Models via Private Cloud Compute at no cloud API cost, lowering the barrier for AI-powered features in smaller apps.
major
Visual Intelligence API for third-party apps
A new VisualIntelligence framework and IntentValueQuery protocol lets apps receive a SemanticContentDescriptor (pixel buffer from camera or screenshot) and return matched app entities directly inside the system Visual Intelligence overlay; the API is now available on macOS and iPadOS in addition to iPhone.
major
MLX Swift — numerical computing and ML framework for Swift
A new MLX Swift package brings NumPy-style N-dimensional array computing to Swift with automatic GPU execution, lazy evaluation, automatic differentiation via grad(), linear algebra, FFTs, convolutions, and optimizers (SGD, Adam, RMSprop), all via Swift Package Manager; companion packages mlx-swift-lm and mlx-swift-examples add LLM inference and Stable Diffusion.
notable
Official Anthropic Claude and Google Gemini Swift packages for Foundation Models
Apple is shipping official Swift packages from Anthropic and Google that plug Claude and Gemini into the Foundation Models LanguageModelSession API via OAuth/Keychain auth, allowing model swapping with a single line of code and per-token usage reporting.
notable
SpotlightSearchTool for Foundation Models — fully local RAG
A new built-in SpotlightSearchTool lets LanguageModelSession autonomously query the device's Core Spotlight index for Retrieval-Augmented Generation without any server calls; GuidanceProfile scopes exposed search capabilities to stay within on-device context limits, and CustomStage enables multi-stage on-device pipelines.
notable
Built-in Vision tools for Foundation Models (BarcodeReaderTool and OCRTool)
New Vision-framework-backed system tools — BarcodeReaderTool and an OCRTool supporting 30+ languages — can be passed directly to LanguageModelSession, letting the on-device model scan barcodes, QR codes, and dense text as part of structured generation tasks.
notable
Foundation Models Python SDK (apple_fm_sdk) and fm CLI tool
A new pip-installable Python SDK (pip install apple_fm_sdk) and a pre-installed fm CLI (fm chat, fm respond, fm schema) let Python developers and shell scripts access Foundation Models on Apple Silicon Macs, including multimodal queries (--image flag), structured JSON output (--schema), and Private Cloud Compute (--model pcc).
notable
Core AI Models open-source repository (github.com/apple/ai-models)
Apple published an open-source repository of curated, Apple-Silicon-ready model weights (including Qwen, Mistral, SAM 3) with conversion and optimization recipes and Swift packages, lowering the barrier to on-device deployment of popular open-source models.
notable
ModelJudgeEvaluator with Cohen's kappa calibration and few-shot examples
ModelJudgeEvaluator uses a language model (including PrivateCloudComputeLanguageModel) to score AI outputs on subjective ScoreDimensions; Cohen's kappa aggregation measures alignment between model judge and human expert ratings, and ModelJudgePrompt supports few-shot worked examples for judge calibration.
notable
LanguageModelExecutor protocol for plugging any LLM into Foundation Models
A new LanguageModelExecutor protocol lets developers package any local or server-based LLM — including community models from Hugging Face MLX-Community — into Foundation Models' LanguageModelSession API; custom Transcript.CustomSegment types can pass structured non-text data (audio, web results) through the pipeline.
notable
Neural Accelerator in M5 GPU shader cores
Apple M5 introduces a dedicated Neural Accelerator block in each GPU shader core, making matrix multiplication four times faster than M4; the Metal Tensor API and MLX automatically target this new hardware, accelerating both custom ML kernels and LLM prompt processing with no code changes.
notable
Updated on-device SystemLanguageModel with improved instruction following, tool calling, and image input
The on-device SystemLanguageModel is rebuilt with better logic, enhanced tool calling, refined guardrails with fewer false positives (iOS 26.4+), native multimodal image input (UIImage, NSImage, CGImage, etc.), and new introspection APIs (model.contextSize, model.tokenCount(for:)).
notable
Reality Composer Pro AI Assistant for 3D content generation
A new AI Assistant panel in Reality Composer Pro 3 generates 3D objects and materials from natural language prompts and integrates Coding Intelligence for generating Swift event-subscription boilerplate (scene.subscribe calls and SwiftUI Attachments) to complement ScriptGraph scenes.
notable
DynamicInstructions and Session Properties for composable agentic context
DynamicInstructions groups tools and instructions into nestable, composable units conditionally included based on app state; the @SessionPropertyEntry macro lets Tools and Profiles share typed state across an agentic graph without external state management.
notable
Open-source Foundation Models utilities Swift package with agentic patterns
A companion open-source Swift package ships community utilities such as rollingWindow(), droppingCompletedToolCalls(), and a Skills API for modular on-demand knowledge injection, updated between OS releases to distribute experimental patterns faster than OS-linked APIs.
notable
MLX-LM Server — OpenAI-compatible HTTP server with continuous batching and distributed inference
MLX-LM Server is an OpenAI-compatible HTTP server exposing local models through a standard API, now supporting structured tool calling, reasoning models, and continuous batching for concurrent multi-agent requests; it also supports distributed inference via mlx.launch to shard large models across multiple Macs.
notable
CoreAIImageSegmenter API wrapping SAM 3 for on-device image segmentation
A new Swift API wrapping the Segment Anything Model 3 (SAM 3) performs text-prompted image segmentation entirely on-device with a simple segment(image:prompt:) call, available via the Core AI framework.
notable
Next-generation photo editing with Apple Intelligence in iOS 27
Apple Intelligence in iOS 27 adds new capabilities to extend, reframe, and clean up photos, going beyond prior generation editing tools.
notable
Use Model action in Shortcuts with web retrieval
The Use Model action in Shortcuts is updated with access to newer Apple Intelligence models that can also retrieve up-to-date information from the web, enabling richer intelligent automation.
notable
LanguageModelCapabilities API and LanguageModelError standardized error taxonomy
Models declare supported capabilities (toolCalling, guidedGeneration, reasoning) via LanguageModelCapabilities so the framework can route requests appropriately; a new LanguageModelError enum defines standard error cases (contextSizeExceeded, rateLimited, refusal, guardrailViolation, timeout, etc.) for consistent error handling across all executor implementations.
notable
historyTransform modifier and TranscriptErrorHandlingPolicy for context engineering
The historyTransform modifier applies stateless per-request transformations to session history before model submission (filtering, trimming) without permanently mutating the transcript; TranscriptErrorHandlingPolicy.preserveTranscript enables mid-response cancellation and resumption.
notable
toolCallingMode API and lifecycle modifiers (onResponse, onToolCall)
A new toolCallingMode modifier (.allowed / .disallowed / .required) gives explicit control over tool calling per prompt, critical for agentic loops; onResponse and onToolCall lifecycle modifiers let developers run imperative code at response boundaries to update UI state or mutate session history.
notable
Visual Intelligence system-level actions (Contacts, Calendar, HealthKit)
Visual Intelligence now supports system-level actions including adding to Contacts, saving calendar events via EventKit, and logging medical device readings via HealthKit, with guidance on observing store-change notifications (e.g., EKEventStoreChanged).
notable
ImageReference and image-based tool calling in Foundation Models
A new ImageReference type and @SessionProperty(..history) allow Tool-protocol-conforming types to receive and resolve image arguments from the session transcript, enabling LLM-orchestrated workflows where the model invokes app-defined image processing functions.
notable
Spotlighting API for prompt injection mitigation
A new probabilistic mitigation technique exposed via .historyTransform wraps untrusted tool output segments with special delimiter tags (e.g., <<UNTRUSTED>>) so the on-device model can recognize and handle external data with appropriate suspicion, reducing prompt injection risk in agentic apps.
notable
JACCL open-source collective communication library for Apple Silicon
Apple released JACCL, an open-source collective communication library built on RDMA over Thunderbolt that provides all-reduce, all-sum, and other primitives for distributed workloads via a C++ API, usable independently of MLX.
notable
TrajectoryExpectation and argument matchers for agentic tool-call evaluation
TrajectoryExpectation lets developers verify the model's tool-call decision path (which tools must be called, in what order, with what arguments, and which must never be called); a rich set of argument matchers (exact, naturalLanguage, contains, oneOf, pattern, range, keyOnly) provides flexible validation, and TrajectoryExpectation is Generable for synthetic dataset generation.
minor
SampleGenerator synthetic data generation with random and slidingWindow strategies
SampleGenerator's async stream API generates evaluation samples from a seed dataset using on-device or cloud models; two built-in sampling strategies control seed example presentation — random for unordered datasets and slidingWindow for datasets with meaningful order.
minor
LanguageModelSession runtime updates and Server-side tool transparency levels
LanguageModelSession now supports continuous runtime updates to models, tools, and instructions; executors can surface server-side tool calls (e.g., web search) at three transparency levels — privately grounded, metadata-enriched, or fully surfaced as custom segments.
minor
save_intermediates API and Core AI Debugger for model conversion debugging
A new coreai API executes a PyTorch model and captures intermediate tensor values at every operation into an Intermediates File, which can be loaded into Core AI Debugger's comparison mode to identify where quantization or conversion diverges from the original.
minor
coreai-opt weight compression with int4/int8/FP4/FP8 presets
The coreai-opt library ships preset compression configurations (e.g., presets.w4 for 4-bit symmetric quantization) and lower-level Quantizer and KMeansPalettizer APIs supporting EAGER and GRAPH execution modes for fine-grained per-layer compression.
minor
MLX_METAL_FAST_SYNCH and nn.layers.distributed APIs
A new MLX_METAL_FAST_SYNCH=1 environment variable enables faster GPU-to-CPU synchronization critical for distributed JACCL communication; new Python APIs shard_linear() and sharded_load() provide fine-grained control over model sharding across a cluster.
minor
ImagePlayground integration with PaperKit
Developers can now launch ImagePlaygroundViewController from a PaperKit adornment tap, receive a generated image URL via delegate, and insert it directly into canvas markup as an ImageMarkup element.
minor
Custom DynamicProfileModifier and open-source orchestration patterns
Developers can conform to DynamicProfileModifier and extend DynamicProfile to create shareable, named modifiers (mirroring SwiftUI's view modifier pattern); baton-pass and phone-a-friend are formally documented as reusable multi-model orchestration blueprints.
minor
Foundation Models framework support extended to Linux
Custom executor Swift packages can now target Linux in addition to iOS, macOS, visionOS, and watchOS, broadening the Foundation Models ecosystem to server-side Swift deployments.