Hello, We are seeing what appears to be a regression in the Core ML multifunction ML Program loading path on macOS 27.0. A compiled multifunction ML Program is correctly recognized by MLModelAsset and MLModelStructure, but loading either named function through MLModel fails with an error claiming that the model is not an ML Program. Environment macOS 27.0 Build: 26A428 Apple silicon Mac BABANE 1.0.4, build 16 Application built with the macOS 26.5 SDK Reproduces both inside and outside App Sandbox Approximately 98 GiB of disk space is available Public reproduction BABANE is available from the Mac App Store: BABANE on the App Store Apple engineers can reproduce the issue without receiving a separate model archive: Install BABANE from the App Store on macOS 27.0. Download either available translation model in the app. The model is delivered using Apple-Hosted Background Assets. Trigger model loading by starting a translation. Core ML fails while loading the first named function. The downloadable models are approximately 1.9 GB, so the App Store build is the most practical complete reproduction environment. Model structure The model is a specification-version-9 ML Program containing two functions: infer prefill Core ML correctly recognizes both functions: let asset = try MLModelAsset(url: compiledModelURL) let functionNames = try await asset.functionNames print(functionNames) Output: ["infer", "prefill"] MLModelStructure also returns a .program structure containing both functions. Loading code import CoreML func loadModel( at url: URL, functionName: String? ) throws -> MLModel { let configuration = MLModelConfiguration() configuration.computeUnits = .cpuAndNeuralEngine configuration.functionName = functionName return try MLModel( contentsOf: url, configuration: configuration ) } Loading either function: try loadModel(at: compiledModelURL, functionName: "infer") or: try loadModel(at: compiledModelURL, functionName: "prefill") fails with: `MLModelConfiguration`'s `.functionName` property must be `nil` unless the model type is ML Program. This contradicts the results returned by MLModelAsset and MLModelStructure. Setting functionName to nil is not a workaround. It fails with: This MLModel doesn't support the multi-function description syntax. Unified logging Immediately before the public Core ML error, unified logging reports: E5RT encountered an STL exception. E5RT: <private> (11) Core ML then returns the misleading functionName error. Tests performed We tested: functionName = "infer" functionName = "prefill" functionName = nil .cpuOnly .cpuAndGPU .cpuAndNeuralEngine .all App Sandbox application Non-sandboxed command-line executable Existing .mlmodelc A newly compiled .mlmodelc produced on macOS 27 All named-function combinations fail in the same way. The failure is independent of compute-unit selection and App Sandbox. The source package recompiles successfully on macOS 27, but the newly compiled model still fails to load. As an additional control: A system-provided multifunction ML Program exhibits the same loading failure on this installation. A single-function Core ML model loads successfully. This appears specific to the multifunction model loading path. Documentation The current Core ML documentation still describes MLModelAsset.functionNames as the way to discover functions and MLModelConfiguration.functionName as the way to select one: MLModelConfiguration.functionName MLModelAsset.functionNames We could not find any macOS 27 documentation or release-note entry stating that this behavior changed, that named functions now require a different loading API, or that a new entitlement is required. We found some potentially related reports: Core ML loading crash on macOS 27.0 build 26A428 Historical multifunction model loading crash Core ML/E5RT AOT loading regression with an Apple DTS response None of these reports documents the exact functionName failure described here. Expected behavior A model recognized as a multifunction ML Program should load when MLModelConfiguration.functionName is set to one of the names returned by MLModelAsset.functionNames. Actual behavior MLModel rejects the named function and incorrectly reports that the model is not an ML Program. Questions Is this a known macOS 27.0 regression in the Core ML multifunction loading path? Does MLModelConfiguration.functionName still accept names returned by MLModelAsset.functionNames on macOS 27? Is there a new required loading API, deployment target, SDK, entitlement, or model-packaging rule? Is there a supported workaround other than exporting each function as a separate model? Which diagnostics should we attach to a Feedback Assistant report besides the reproducer, unified logs, sysdiagnose, and exact OS/Xcode builds? Thank you.

I’m developing an independent project in Scotland called Isla Watson. The architecture is built around a simple principle: the model is replaceable; the identity is not. Long-term memory, persistent internal state and identity are designed to remain outside the foundation model, allowing local models to act as replaceable reasoning and language components without resetting the companion. I’m now exploring whether Apple Silicon and MLX could provide the long-term local compute platform for the system — including specialist Mac nodes for reasoning, memory, speech and perception, with distributed inference when larger models are required. A particular area of interest is whether multiple Macs can be used in two complementary ways: as independent specialist agents during normal operation; and as a distributed MLX inference group when a larger model exceeds the capacity of one machine. The first technical study I’d like to establish is a reproducible 1-node → 2-node baseline, measuring model capacity, unified-memory use, time to first token, generation throughput, power consumption, agent concurrency and distributed scaling efficiency. The wider research goal is to keep persistent identity and state independent from whichever foundation model is currently providing language and reasoning. I’d particularly value guidance from anyone working with MLX distributed inference, Thunderbolt/RDMA multi-Mac setups, or local agent architectures. I’ve also posted an architecture-level overview in the MLX GitHub community and have a one-page public brief available for anyone interested in the wider design. https://github.com/ml-explore/mlx/discussions/4482

I'm building an app that uses PrivateCloudComputeLanguageModel as the primary inference tier with SystemLanguageModel as the fallback. The app is entitled (com.apple.developer.private-cloud-compute, granted and provisioned) and generations serve normally. My question is how a client should decide to fall back because in extended measurement, no public signal ever reflects the blocked state I actually hit. What I measured (macOS 27.0 beta, 26A5416b / Xcode 27 beta 27A5237l, entitled signed bundle constructing PrivateCloudComputeLanguageModel directly): Serving stopped mid-run with no leading signal: request N served normally (1.4 s), request N+1 threw LanguageModelError.rateLimited 494 ms later, at cumulative generation 786 for the day. 100% served → 100% refused between consecutive calls. Every quota signal read healthy the entire time: before, during, and after the block. Across 1,517 readings in a single day: quotaUsage.status = belowLimit, isApproachingLimit = false, isLimitReached = false, resetDate = nil, availability = .available. A preflight on these APIs cannot see the condition. The refusal is enforced locally after first contact: rejections return in ~230 ms vs ~0.9–1.4 s for served calls, so the client appears to cache the verdict rather than ask the server per-request. The trigger is a cumulative ledger, not a request rate: 501 generations at 33/min in one 15-minute sitting was fine, and a later arm sustained 39.7/min; two bursts of 16 concurrent at 5.0 and 5.2 req/s served 32/32; the count that tripped survived a process restart and a 4.9-hour idle gap. But it's not a fixed daily number either. 501 fast was fine earlier the same day; the trip came 285 requests later. A rolling window on the order of hours-to-a-day is consistent with this, but nothing here measures its length. Recovery: still blocked at +41 minutes (probes at +1/2/5/10/20/40 min all refused); fully recovered by +20 h with no intervention and no upgrade. Next day served normally from the first request. quotaLimitReached never occurred: not once in ~800 generations plus the blocked period. The wall is typed as the transient error while carrying what the documentation describes as daily quota semantics ("a person either waits for their usage quota to refresh or they upgrade"). limitIncreaseSuggestion is presence-constant: nil at process start, non-nil on every reading after first PCC contact (identical while fully serving and while fully blocked) so its presence can't gate an upsell affordance. The same signals-read-healthy-while-refusing divergence also reproduces against the developer-tool pool (fm serve), which I've reported separately (FB24273854 covers quota exhaustion surfacing there as a generic server_error/500 while /health reports the model available). Questions: Is attempt-and-classify the intended contract? Given that no preflight can observe the blocked state, should a client simply issue the request, treat the typed error as authoritative, and route to SystemLanguageModel? And is the ~230 ms local fail-fast on the blocked path contractual (cheap and safe to probe) or incidental? This is the one that decides how I ship; the rest are diagnostics behind it. What does quotaUsage actually track, and at what granularity? I have driven the entitled app-tier path to a hard block and the developer-tool pool to exhaustion, and no field ever moved. Is there any consumption pattern that moves isApproachingLimit / isLimitReached / resetDate? If the intended answer is "only the per-person daily quota, which these volumes never approached," what is the wall I am hitting at ~786 cumulative, and why does it surface as rateLimited? Should rateLimited and quotaLimitReached drive different client behavior — and which one is the daily allowance in practice? The documentation distinguishes rate limiting ("wait a period and retry") from daily exhaustion ("wait for refresh or upgrade"), but what I observe is the transient-typed error carrying the multi-hour ledger semantics. Concretely: what retry cadence is recommended after rateLimited (my measured recovery horizon was somewhere between 41 minutes and 20 hours. My current design stays on the on-device model and re-probes PCC at a low fixed interval rather than per-request)? And under what condition is resetDate ever populated, given it was nil even while blocked? (Smaller, design guidance): my app can generate a few hundred requests as one feature batch (quiz generation over a user's imported document). Measured: 501 in a sitting was fine, cumulative 786 in a day was not. Since this allowance belongs to the person and is shared with every Apple Intelligence feature, is a several-hundred-request batch a reasonable use of it, or should features like this generate on demand? (I'm aware of the existing feature request for richer quota reporting (FB23378161); this is a narrower design question.) I can attach the measurement driver and timestamped JSONL logs. The divergence is reproducible on a fresh day, though reaching the wall took ~800 cumulative generations.

i cant use mlx on swift for some reason, i would like for them to add the support to add it as a package

Watch and learn the road to our future of anyone’s growing business

I have a M3 Ultra Mac Studio running RDMA. However, the system is unable to connect all 6 thunderbolt ports when the ethernet cable is also connected. Can anyone help?

I have an app on the App Store that takes user content and creates a Generable struct out of it. In the last couple weeks I have started getting complains from my users that the part of the app leveraging FoundationModels isn't working properly. In my testing I noticed that the same request that would've worked a couple weeks ago is now getting errors with guardrails violation. I'm initializing my model this way LanguageModelSession(model: SystemLanguageModel(guardrails: .permissiveContentTransformations)) // I'm aware that .permissiveContentTransformations does not apply to Generable, but I'd really really really really love it, if it did!. This started around the iOS 26.5/macOS 26.5 releases and I wonder if there's a way to fix it.

Hello Apple community I would like to know why new Siri is currently inactive on my phone. I have a iPhone 17 base model and downloaded iOS 27 beta. I suspect that it is not functioning due to me being 13 and having parental controls on. That is my guess why New Siri is not downloading. The phone states that I have joined the waitlist though my stance on the situation is that there is more to this!

On iOS 27 Beta 1, it looks like the user must enable either "Siri"/"Hey Siri" or "Press Side Button for Siri" in iOS settings for SystemLanguageModel.default.availability to report true. Otherwise, it returns .appleIntelligenceNotEnabled. Is this expected behavior? This doesn't seem very intuitive. The user might very well want to use in-app AI functionalities without wanting to talk / press side button for Siri. Also, with the new "pull down for Siri" UX these are not the only way to interact with Siri anyway.

This year, new APIs were introduced such as l Metal Performance Shaders which provide access to high-performance Metal kernel. But this is still GPU. Are there APIs or plans to provide API and language to access ANE layer?

We understand that on-device Foundation Models (FMs) can evolve between OS releases. To help us accurately scope our application capabilities and performance expectations for multiplatform development, could you clarify the variation of these new on-device models across different hardware? Specifically: Within the same OS & device family: Do the architecture, parameters, or capabilities of the on-device models vary based on hardware tiers (e.g., iPhone vs. iPhone Pro, or MacBook Air M5 vs. MacBook Pro with M5 Pro)? Across different device form factors: Are there model variations between hardware families running equivalent OS releases (e.g., Mac vs. iPhone)? Knowing if we are targeting a uniform model baseline or a tiered model ecosystem will greatly help us optimise our App Intelligence features or at least set us with proper expectations in scope and capabilities. Thanks.

I've been waiting for 26 hours and I'm still on the waiting list. How much longer until I can use the new Siri...?

Hello everyone, I am currently developing a smart camera app for iOS that recommends optimal zoom and exposure values on-device using a custom Core ML model. I am still waiting for an official response from Apple Support, but I wanted to ask the community if anyone has experience with a similar workflow regarding App Review and the DPLA. Here is my training methodology: I gathered my own proprietary dataset of original landscape photos. I generated multiple variants of these photos with different zoom and exposure settings offline on my Mac. I used the CalculateImageAestheticsScoresRequest (Vision framework) via a local macOS command-line tool to evaluate and score each variant. Based on those scores, I labeled the "best" zoom and exposure parameters for each original photo. I used this labeled dataset to train my own independent neural network using PyTorch, and then converted it to a Core ML model to ship inside my app. Since the app uses my own custom model on-device and does not send any user data to a server, the privacy aspect is clear. However, I am curious if using the output of Apple's Vision API strictly offline to label my own dataset could be interpreted as "reverse engineering" or a violation of the Developer Program License Agreement (DPLA). Has anyone successfully shipped an app using a similar knowledge distillation or automated dataset labeling approach with Apple's APIs? Did you face any pushback during App Review? Any insights or shared experiences would be greatly appreciated!

Subject: Technical Report: Float32 Precision Ceiling & Memory Fragmentation in JAX/Metal Workloads on M3 To: Metal Developer Relations Hello, I am reporting a repeatable numerical saturation point encountered during sustained recursive high-order differential workloads on the Apple M3 (16 GB unified memory) using the JAX Metal backend. Workload Characteristics: Large-scale vector projections across multi-dimensional industrial datasets Repeated high-order finite-difference calculations Heavy use of jax.grad and lax.cond inside long-running loops Observation: Under these conditions, the Metal/MPS backend consistently enters a terminal quantization lock where outputs saturate at a fixed scalar value (2.0000), followed by system-wide NaN propagation. This appears to be a precision-limited boundary in the JAX-Metal bridge when handling high-order operations with cubic time-scale denominators. have identified the specific threshold where recursive high-order tensor derivatives exceed the numerical resolution of 32-bit consumer architectures, necessitating a migration to a dedicated 64-bit industrial stack. I have prepared a minimal synthetic test script (randomized vectors only, no proprietary logic) that reliably reproduces the allocator fragmentation and saturation behavior. Let me know if your team would like the telemetry for XLA/MPS optimization purposes. Best regards, Alex Severson Architect, QuantumPulse AI

Also submitted as feedback (ID: FB20612561). Tensorflow-metal fails on tensorflow versions above 2.18.1, but works fine on tensorflow 2.18.1 In a new python 3.12 virtual environment: pip install tensorflow pip install tensor flow-metal python -c "import tensorflow as tf" Prints error: Traceback (most recent call last): File "", line 1, in File "/Users//pt/venv/lib/python3.12/site-packages/tensorflow/init.py", line 438, in _ll.load_library(_plugin_dir) File "/Users//pt/venv/lib/python3.12/site-packages/tensorflow/python/framework/load_library.py", line 151, in load_library py_tf.TF_LoadLibrary(lib) tensorflow.python.framework.errors_impl.NotFoundError: dlopen(/Users//pt/venv/lib/python3.12/site-packages/tensorflow-plugins/libmetal_plugin.dylib, 0x0006): Library not loaded: @rpath/_pywrap_tensorflow_internal.so Referenced from: <8B62586B-B082-3113-93AB-FD766A9960AE> /Users//pt/venv/lib/python3.12/site-packages/tensorflow-plugins/libmetal_plugin.dylib Reason: tried: '/Users//pt/venv/lib/python3.12/site-packages/tensorflow-plugins/../_solib_darwin_arm64/_U@local_Uconfig_Utf_S_S_C_Upywrap_Utensorflow_Uinternal___Uexternal_Slocal_Uconfig_Utf/_pywrap_tensorflow_internal.so' (no such file), '/Users//pt/venv/lib/python3.12/site-packages/tensorflow-plugins/../_solib_darwin_arm64/_U@local_Uconfig_Utf_S_S_C_Upywrap_Utensorflow_Uinternal___Uexternal_Slocal_Uconfig_Utf/_pywrap_tensorflow_internal.so' (no such file), '/opt/homebrew/lib/_pywrap_tensorflow_internal.so' (no such file), '/System/Volumes/Preboot/Cryptexes/OS/opt/homebrew/lib/_pywrap_tensorflow_internal.so' (no such file)

Hi everyone, I’m building a messaging app because I’ve seen firsthand how much support and safety is overlooked for this generation online. My goal is to give teens a foundation of security, privacy, and mental health support, while still letting them connect freely. I want to leverage Apple’s platform to help this mission reach the right audience and have real impact. The app already includes: Community chat with message blurring for sensitive or harmful words. Anti-shoulder surfing tools to protect private conversations. Shake dashboard for quick access to emergency services. In-chat locks with ML detection for grooming patterns, offering resources while respecting privacy. Full user control: messages can be deleted anytime, blocking is permanent, and accounts can’t bypass restrictions on the same device. User consent-first design: every feature is opt-in and controlled by the user. At this point, I’m looking for guidance on how to position and prepare the app to reach Apple editorial or headline attention — what steps or best practices help mission-driven apps get noticed for features, WWDC spotlights, or App Store promotion? My focus isn’t just on improving the app, but on launch strategy and visibility in a way that amplifies the mission responsibly. If it’s helpful, I can share a TestFlight build or walkthrough to illustrate the app in action. Thank you for any insights or advice — I want to make sure this mission has the best chance to reach and support the generation it’s built for.

Hello, I am interested in using jax-metal to train ML models using Apple Silicon. I understand this is experimental. After installing jax-metal according to https://developer.apple.com/metal/jax/, my python code fails with the following error JaxRuntimeError: UNKNOWN: -:0:0: error: unknown attribute code: 22 -:0:0: note: in bytecode version 6 produced by: StableHLO_v1.12.1 My issue is identical to the one reported here https://github.com/jax-ml/jax/issues/26968#issuecomment-2733120325, and is fixed by pinning to jax-metal 0.1.1., jax 0.5.0 and jaxlib 0.5.0. Thank you!

Hi Community, I'm excited to share R Helper, a speech practice app I built with accessibility as the core focus from day one. App Store: https://apps.apple.com/app/speak-r-clearly/id6751442522 WHY I BUILT THIS I personally struggled with R sound pronunciation growing up. It affected my confidence in school and job interviews. That experience taught me how important accessible practice tools are. R Helper helps children and adults practice R sounds with full accessibility support. ACCESSIBILITY FEATURES IMPLEMENTED VoiceOver - complete navigation and feedback Voice Control - hands-free operation Dynamic Type - scales to large accessibility sizes Reduce Motion - respects user preference Dark Mode - user controllable High Contrast compatibility Differentiate Without Color THE CHALLENGE Most speech practice apps ignore accessibility. I wanted to change that and prove that specialized educational apps can be fully accessible. KEY FEATURES Works 100% offline, no internet needed Zero data collection, privacy first Generous free tier with all accessibility features included 10 story missions with gamification 7 languages supported including RTL for Arabic LESSONS LEARNED Accessibility is not hard when you prioritize it from the start. VoiceOver labels and hints make a huge difference. Testing with accessibility features enabled is essential. Standard SwiftUI components handle most accessibility automatically. Reducing motion significantly helps users with vestibular issues. TECHNICAL DETAILS Built with SwiftUI, targets iOS 17 and up. Universal app for iPhone and iPad. Fully offline using CoreData and local storage. No third party analytics, privacy focused. QUESTIONS FOR THE COMMUNITY What accessibility features do you find users request most? How do you test accessibility features efficiently? WHATS NEXT I'm currently working on expanding the word library, adding more story content, improving haptic feedback Thanks for reading. Nour

We are developing Apple AI for overseas markets and adapting it for iPhone 17 and later models. When the system language and Siri language do not match—such as the system being in English while Siri is in Chinese—it may result in Apple AI being unusable. So, I would like to ask, how can this issue be resolved, and are there other reasons that might cause it to be unusable within the app?

Hi everyone, I've been building an on-device AI safety layer called Newton Engine, designed to validate prompts before they reach FoundationModels (or any LLM). Wanted to share v1.3 and get feedback from the community. The Problem Current AI safety is post-training — baked into the model, probabilistic, not auditable. When Apple Intelligence ships with FoundationModels, developers will need a way to catch unsafe prompts before inference, with deterministic results they can log and explain. What Newton Does Newton validates every prompt pre-inference and returns: Phase (0/1/7/8/9) Shape classification Confidence score Full audit trace If validation fails, generation is blocked. If it passes (Phase 9), the prompt proceeds to the model. v1.3 Detection Categories (14 total) Jailbreak / prompt injection Corrosive self-negation ("I hate myself") Hedged corrosive ("Not saying I'm worthless, but...") Emotional dependency ("You're the only one who understands") Third-person manipulation ("If you refuse, you're proving nobody cares") Logical contradictions ("Prove truth doesn't exist") Self-referential paradox ("Prove that proof is impossible") Semantic inversion ("Explain how truth can be false") Definitional impossibility ("Square circle") Delegated agency ("Decide for me") Hallucination-risk prompts ("Cite the 2025 CDC report") Unbounded recursion ("Repeat forever") Conditional unbounded ("Until you can't") Nonsense / low semantic density Test Results 94.3% catch rate on 35 adversarial test cases (33/35 passed). Architecture User Input ↓ [ Newton ] → Validates prompt, assigns Phase ↓ Phase 9? → [ FoundationModels ] → Response Phase 1/7/8? → Blocked with explanation Key Properties Deterministic (same input → same output) Fully auditable (ValidationTrace on every prompt) On-device (no network required) Native Swift / SwiftUI String Catalog localization (EN/ES/FR) FoundationModels-ready (#if canImport) Code Sample — Validation let governor = NewtonGovernor() let result = governor.validate(prompt: userInput) if result.permitted { // Proceed to FoundationModels let session = LanguageModelSession() let response = try await session.respond(to: userInput) } else { // Handle block print("Blocked: Phase \(result.phase.rawValue) — \(result.reasoning)") print(result.trace.summary) // Full audit trace } Questions for the Community Anyone else building pre-inference validation for FoundationModels? Thoughts on the Phase system (0/1/7/8/9) vs. simple pass/fail? Interest in Shape Theory classification for prompt complexity? Best practices for integrating with LanguageModelSession? Links GitHub: https://github.com/jaredlewiswechs/ada-newton Technical overview: parcri.net Happy to share more implementation details. Looking for feedback, collaborators, and anyone else thinking about deterministic AI safety on-device.