Summary Since macOS 26, our Core ML / MPS inference pipeline produces incorrect results on Mac mini M1 (Macmini9,1, A14-class SoC). The same model and code runs correctly on M2 and newer (A15-class and up). The regression appears to be in the Scaled Dot-Product Attention (SDPA) kernel path in the MPS backend. Environment Affected Mac mini M1 — Macmini9,1 (A14-class) Not affected M2 and newer (A15-class and up) Last known good macOS Sequoia First broken macOS 26 (Tahoe) ? Confirmed broken on macOS 26.3.1 Framework Core ML + MPS backend Language C++ (via CoreML C++ API) Description We ship an audio processing application (VoiceAssist by NoiseWorks) that runs a deep learning model (based on Demucs architecture) via Core ML with the MPS compute unit. On macOS Sequoia this works correctly on all Apple Silicon Macs including M1. After updating to macOS 26 (Tahoe), inference on M1 Macs fails — either producing garbage output or crashing. The same binary, same .mlpackage, same inputs work correctly on M2+. Our Apple contact has suggested the root cause is a regression in the A14-specific MPS SDPA attention kernel, which may have broken when the Metal/MPS stack was updated in macOS 26. The model makes heavy use of attention layers, and the failure correlates precisely with the SDPA path being exercised on A14 hardware. Steps to Reproduce Load a Core ML model that uses Scaled Dot-Product Attention (e.g. a transformer or attention-based audio model) Run inference with MLComputeUnits::cpuAndGPU (MPS active) Run on Mac mini M1 (Macmini9,1) with macOS 26.3.1 Compare output to the same model running on M2 / macOS Sequoia Expected: Correct inference output, consistent with M2+ and macOS Sequoia behavior Actual: Incorrect / corrupted output (or crash), only on A14-class hardware running macOS 26+ Workaround Forcing MLComputeUnits::cpuOnly bypasses MPS entirely and produces correct output on M1, confirming the issue is in the MPS compute path. This is not acceptable as a shipping workaround due to performance impact. Additional Notes The failure is hardware-specific (A14 only) and OS-specific (macOS 26+), pointing to a kernel-level regression rather than a model or app bug We first became aware of this through a customer report Happy to provide a symbolicated crash log if helpful this text was summarized by AI and human verified

Hello, I’m experiencing a severe performance degradation when running CoreML models on a live AVFoundation video feed compared to offline or synthetic inference. This happens across multiple models I've converted (including SCI, RTMPose, and RTMW) and affects multiple devices. The Environment OS: macOS 26.3, iOS 26.3, iPadOS 26.3 Hardware: Mac14,6 (M2 Max), iPad Pro 11 M1, iPhone 13 mini Compute Units: cpuAndNeuralEngine The Numbers When testing my SCI_output_image_int8.mlpackage model, the inference timings are drastically different: Synthetic/Offline Inference: ~1.34 ms Live Camera Inference: ~15.96 ms Preprocessing is completely ruled out as the bottleneck. My profiling shows total preprocessing (nearest-neighbor resize + feature provider creation) takes only ~0.4 ms in camera mode. Furthermore, no frames are being dropped. What I've Tried I am building a latency-critical app and have implemented almost every recommended optimization to try and fix this, but the camera-feed penalty remains: Matched the AVFoundation camera output format exactly to the model input (640x480 at 30/60fps). Used IOSurface-backed pixel buffers for everything (camera output, synthetic buffer, and resize buffer). Enabled outputBackings. Loaded the model once and reused it for all predictions. Configured MLModelConfiguration with reshapeFrequency = .frequent and specializationStrategy = .fastPrediction. Wrapped inference in ProcessInfo.processInfo.beginActivity(options: .latencyCritical, reason: "CoreML_Inference"). Set DispatchQueue to qos: .userInteractive. Disabled the idle timer and enabled iOS Game Mode. Exported models using coremltools 9.0 (deployment target iOS 26) with ImageType inputs/outputs and INT8 quantization. Reproduction To completely rule out UI or rendering overhead, I wrote a standalone Swift CLI script that isolates the AVFoundation and CoreML pipeline. The script clearly demonstrates the ~15ms latency on live camera frames versus the ~1ms latency on synthetic buffers. (I have attached camera_coreml_benchmark.swift and coreml model (very light low light enghancement model) to this repo on github https://github.com/pzoltowski/apple-coreml-camera-latency-repro). My Question: Is this massive overhead expected behavior for AVFoundation + Core ML on live feeds, or is this a framework/runtime bug? If expected, what is the Apple-recommended pattern to bypass this camera-only inference slowdown? One think found interesting when running in debug model was faster (not as fast as in performance benchmark but faster than 16ms. Also somehow if I did some dummy calculation on on different DispatchQueue also seems like model got slightly faster. So maybe its related to ANE Power State issues (Jitter/SoC Wake) and going to fast to sleep and taking a long time to wakeup? Doing dummy calculation in background thought is probably not a solution. Thanks in advance for any insights!

Hi, I'm building ANF (Autonomous Native Forge) — a cloud-free, 4-agent autonomous software production pipeline running on local hardware with local LLM inference. No middleware, pure Node.js native. Currently running on NVIDIA Blackwell GB10 with vLLM + DeepSeek-R1-32B. Now porting to Apple Silicon. Three technical questions: How production-ready is mlx-lm's OpenAI-compatible API server for long context generation (32K tokens)? What's the recommended approach for KV Cache management with Unified Memory architecture — any specific flags or configurations for M4 Ultra? MLX vs GGUF (llama.cpp) for a multi-agent pipeline where 4 agents call the inference endpoint concurrently — which handles parallel requests better on Apple Silicon? GitHub: github.com/trgysvc/AutonomousNativeForge Any guidance appreciated.

Environment: Hardware: Mac M4 OS: macOS Sequoia 15.7.4 TensorFlow-macOS Version: 2.16.2 TensorFlow-metal Version: 1.2.0 Description: When using the tensorflow-metal plug-in for GPU acceleration on M4, the ReLU activation function (both as a layer and as an activation argument) fails to correctly clip negative values to zero. The same code works correctly when forced to run on the CPU. Reproduction Script: import os import numpy as np import tensorflow as tf # weights and biases = -1 weights = [np.ones((10, 5)) * -1, np.ones(5) * -1] # input = 1 data = np.ones((1, 10)) # comment this line => GPU => get negative values # uncomment this line => CPU => no negative values # tf.config.set_visible_devices([], 'GPU') # create model model = tf.keras.Sequential([ tf.keras.layers.Input(shape=(10,)), tf.keras.layers.Dense(5, activation='relu') ]) # set weights model.layers[0].set_weights(weights) # get output output = model.predict(data) # check if negative is present print(f"min value: {output.min()}") print(f"is negative present? {np.any(output < 0)}")

Looking for help with or to help with, due to the pending document enhancement, the Vibe Coders edition of cml editor. Also for more information on how to use the .mlkey whether or not my model is suppose to say IOs18 when I am planning to use it on Mac Apple Intelligence seems to think coreML is for iOS but are the capabilities extended when running NPU on the book? How to use this graph. coming in hot sorry. btw. there are 100s of feedback and crash reports sent in form me for additional info? I attached a image that might help with updating Tags

After exerting a custom model with nms=True. In Xcode, the outputs show as: confidence: MultiArray (0 × 5) coordinates: MultiArray (0 × 4) I want to set fixed shapes (e.g., 100 × 5, 100 × 4), but Xcode does not allow editing—the shape fields are locked. The model graph shows both outputs come directly from a NonMaximumSuppression layer. Is it possible to set fixed output dimensions for NMS outputs in CoreML?

I am trying to benchmark and see if the Qwen3 1.7B model can run in an iPhone SE 3 [4 GB RAM]. My core problem is - Even with weight quantization the SE 3 is not able to load into memory. What I've tried: I am converting a Torch model to the Core ML format using coremltools. I have tried the following combinations of quantization and context length 8 bit + 1024 8 bit + 2048 4 bit + 1024 4 bit + 2048 All the above quantizations are done with dynamic shape with the default being [1,1] in the hope that the whole context length does not get allocated in memory The 4-bit model is approximately 865MB on disk The 8-bit model is approximately 1.7 GB on disk During load: With the int4 quantization the memory spikes during intitial load a lot. Could this be because many operations are converted to int8 or fp16 as core ML does not perform operations natively on int4? With int8 on the profiler the memory does not go above 2 GB (only 900 MB) but it is still not able to load as it shows the following error. 2GB is the limit where jetsam kills the app for the iPhone SE 3 E5RT: Error(s) occurred compiling MIL to BNNS graph: [CreateBnnsGraphProgramFromMIL]: BNNS Graph Compile: failed to preallocate file with error: No space left on device for path: /var/mobile/Containers/Data/Application/ 5B8BB7D2-06A6-4BAE-A042-407B6D805E7C/Library/Caches /com.tss.qwen3-coreml/ com.apple.e5rt.e5bundlecache/ 23A341/<long key>.tmp.12586_4362093968.bundle/ H14.bundle/main/main_bnns/bnns_program.bnnsir Some online sources have suggested activation quantization but I am unsure if that will have any impact on loading [as the spike is during load and not inference] The model spec also suggests that there is no dequantization happening (for e.g from 4 bit -> fp16) So I had couple of queries: Has anyone faced similar issues? What could be the reasons for the temporary memory spike during LOAD What are approaches that can be adopted to deal with this issue? Any help would be greatly appreciated. Thank you.

Optimal Precision • Current Precision: Mixed (Float32, int32) • Optimal Precision: Not specified in the image, but typically involves using the most efficient data type for the model's operations to balance speed and memory usage without significant loss of accuracy. Comparison: • Mixed Precision: Utilizes both Float32 and int32 to optimize performance. Float32 provides high precision, while int32 reduces memory usage and increases computational speed. • Optimal Precision: Aimed at achieving the best trade-off between performance and accuracy, potentially using other data types like Float16 (bfloat16) for even greater efficiency in certain hardware environments. Operation Distribution • Current Distribution: • iOS18.mul: 168 • iOS18.transpose: 126 • iOS18.linear: 98 • iOS18.add: 97 • iOS18.sliceByIndex: 96 • iOS18.expandDims: 74 • iOS18.concat: 72 • iOS18.squeeze: 72 • iOS18.reshape: 67 • iOS18.layerNorm: 49 • iOS18.matmul: 48 • iOS18.gelu: 26 • iOS18.softmax: 24 • Split: 24 • conv: 1 • iOS18.conv: 1 Comparison: • Operation Count: Indicates how frequently each operation is executed. High counts for operations like mul, transpose, and linear suggest these are computationally intensive parts of the model. • Optimization Opportunities: Reducing the count of high-frequency operations or optimizing their execution can improve performance. This might involve pruning unnecessary operations, optimizing algorithms, or leveraging hardware acceleration. General Recommendations • Precision Tuning: Experiment with different precision levels to find the best balance for your specific hardware and accuracy requirements. • Operation Optimization: Focus on optimizing the most frequent operations. Techniques include using more efficient algorithms, parallelizing computations, or utilizing specialized hardware like GPUs or TPUs. • Benchmarking: Regularly benchmark the model to assess the impact of changes and ensure that optimizations lead to meaningful performance improvements. By focusing on these areas, you can potentially enhance the efficiency and performance of your ML model.

Hi all, I spent the last few months developing an MLX/Ollama local AI Benchmarking suite for Apple Silicon, written in pure Swift and signed with an Apple Developer Certificate, open source, GPL, and free. I would love some feedback to continue development. It is the only benchmarking suite I know of that supports live power metrics and MLX natively, as well as quick exports for benchmark results, and an arena mode, Model A vs B with history. I really want this project to succeed, and have widespread use, so getting 75 stars on the github repo makes it eligible for Homebrew/Cask distribution. Github Repo

We are trying to implement a custom encryption scheme for our Core ML models. Our goal is to bundle encrypted models, decrypt them into memory at runtime, and instantiate the MLModel without the unencrypted model file ever touching the disk. We have looked into the native apple encryption described here https://developer.apple.com/documentation/coreml/encrypting-a-model-in-your-app but it has limitations like not working on intel macs, without SIP, and doesn’t work loading from dylib. It seems like most of the Core ML APIs require a file path, there is MLModelAsset APIs but I think they just write a modelc back to disk when compiling but can’t find any information confirming that (also concerned that this seems to be an older API, and means we need to compile at runtime). I am aware that the native encryption will be much more secure but would like not to have the models in readable text on disk. Does anyone know if this is possible or any alternatives to try to obfuscate the Core ML models, thanks

Problem: CoreML produces NaN on GPU (works fine on CPU) when running transformer attention with fused QKV projection on macOS 26.2. Root cause: The common::fuse_transpose_matmul optimization pass triggers a Metal kernel bug when sliced tensors feed into matmul(transpose_y=True). Workaround: pipeline = ct.PassPipeline.DEFAULT pipeline.remove_passes(['common::fuse_transpose_matmul']) mlmodel = ct.convert(model, ..., pass_pipeline=pipeline) Minimal repro: https://github.com/imperatormk/coreml-birefnet/blob/main/apple_bug_repro.py Affected: Any ViT/Swin/transformer with fused QKV attention (BiRefNet, etc.) Has anyone else hit this? Filed FB report too.

About the Core ML model encryption mention in:https://developer.apple.com/documentation/coreml/encrypting-a-model-in-your-app Ｗhen I encrypted the model, if the machine is M chip, the model will load perfectly. One the other hand, when I test the executable on an Intel chip macbook, there will be an error: Error Domain=com.apple.CoreML Code=9 "Operation not supported on this platform." UserInfo={NSLocalizedDescription=Operation not supported on this platform.} Intel test machine is 2019 macbook air with CPU: Intel i5-8210Y, OS: 14.7.6 23H626, With Apple T2 Security Chip. The encrypted model do load on M2 and M4 macbook air. If the model is NOT encrypted, it will also load on the Intel test machine. I did not find in Core ML document that suggest if the encryption/decryption support Intel chips. May I check if the decryption indeed does NOT support Intel chip?

Hello everyone, I’m currently working with the Message Filtering Extension and would really appreciate some clarification around its performance and operational constraints. While the extension is extremely powerful and useful, I’ve found that some important details are either unclear or not well covered in the available documentation. There are two main areas I’m trying to understand better: Machine learning model constraints within the extension In our case, we already have an existing ML model that classifies messages (and are not dependant on Apple's built-in models). We’re evaluating whether and how it can be used inside the extension. Specifically, I’m trying to understand: Are there documented limits on the size of an ML model (e.g., maximum bundle size or model file size in MB)? What are the memory constraints for a model once loaded into memory by the extension? Under what conditions would the system terminate or “kick out” the extension due to memory or performance pressure? Message processing timeouts and execution constraints What is the timeout for processing a single received message? At what point will the OS stop waiting for the extension’s response and allow the message by default (for example, if the extension does not respond in time)? Any guidance, official references, or practical experience from Apple engineers or other developers would be greatly appreciated. Thanks in advance for your help,

We’ve encountered what appears to be a CoreML regression between macOS 26.0.1 and macOS 26.1 Beta. In macOS 26.0.1, CoreML models run and produce correct results. However, in macOS 26.1 Beta, the same models produce scrambled or corrupted outputs, suggesting that tensor memory is being read or written incorrectly. The behavior is consistent with a low-level stride or pointer arithmetic issue — for example, using 16-bit strides on 32-bit data or other mismatches in tensor layout handling. Reproduction Install ON1 Photo RAW 2026 or ON1 Resize 2026 on macOS 26.0.1. Use the newest Highest Quality resize model, which is Stable Diffusion–based and runs through CoreML. Observe correct, high-quality results. Upgrade to macOS 26.1 Beta and run the same operation again. The output becomes visually scrambled or corrupted. We are also seeing similar issues with another Stable Diffusion UNet model that previously worked correctly on macOS 26.0.1. This suggests the regression may affect multiple diffusion-style architectures, likely due to a change in CoreML’s tensor stride, layout computation, or memory alignment between these versions. Notes The affected models are exported using standard CoreML conversion pipelines. No custom operators or third-party CoreML runtime layers are used. The issue reproduces consistently across multiple machines. It would be helpful to know if there were changes to CoreML’s tensor layout, precision handling, or MLCompute backend between macOS 26.0.1 and 26.1 Beta, or if this is a known regression in the current beta.

I'm working on my first model that detects bowling score screens, and I have it working with pictures no problem. But when it comes to video, I have a sizing issue. I added my model to a small app I wrote for taking a picture of a Bowling Scoring Screen, where my model will frame the screens in the video feed from the camera. My model works, but my boxes are about 2/3 the size of the screens being detected. I don't understand the theory of the video stream the camera is feeding me. What I mean is that I don't want to make tweaks to the size of my rectangles by making them larger, and I'm not sure if the video feed is larger than what I'm detecting in code. Questions I have are like is the video feed a certain resolution like 1980x something, or a much higher resolution in the 12 megapixel range? On a static image of say 1920x something, My alignment is perfect. AI says that it's my model training, that I'm training on square images but video is 16:9. Or that I'm producing 4:3 images in a 16:9 environment. I'm missing something here but not sure what it is. I already wrote code to force it to fit, but reverted back to trying for a natural fit.

When trying to open an encrypted CoreML model file on a system with SIP disabled, the error message is Failed to generate key request for <...> with error: -42187 This should state that SIP is disabled and needs to be enabled.

Greetings, and Happy Holidays, 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. parcri.net has the link :)

I’m seeing consistent failures using SoundAnalysis live classification when my app moves to the background. Setup iOS 17.x AVAudioEngine mic capture SNAudioStreamAnalyzer SNClassifySoundRequest(classifierIdentifier: .version1) UIBackgroundModes = audio AVAudioSession .record / .playAndRecord, active Audio capture + level metering continue working in background (mic indicator stays on) Issue As soon as the app enters background / screen locks: SoundAnalysis starts failing every second with domain:com.apple.SoundAnalysis, code:2(SNErrorCode.operationFailed) Audio capture itself continues normally When the app returns to foreground, classification immediately resumes without restarting the engine/analyzer Question Is live background sound classification with the built-in SoundAnalysis classifier officially unsupported or known to fail in background? If so, is a custom Core ML model the only supported approach for background detection? Or is there a required configuration I’m missing to keep SNClassifySoundRequest(.version1) running in background? Thanks for any clarification.

Hi Apple Engineers, I am experiencing a potential memory management bug with CoreML on M1 Mac (32GB Unified Memory). When processing long video files (approx. 12,000 frames) using a CoreML execution provider, the system often completes the 'Analysing' phase but fails to transition into 'Processing'. It simply exits silently or hits an import error (scipy). However, if I split the same task into small 20-frame segments, it works perfectly at high speeds (~40 FPS). This suggests the hardware is capable, but there is an issue with memory fragmentation or resource cleanup during long-running CoreML sessions. Is there a way to force a VRAM/Unified Memory flush via CLI, or is this a known limitation for large frame indexing?

Hi everyone 👋 I'd like to use coremltools to see how well a model performs on a remote device as part of a CI/CD pipeline. According to the Core ML Tools "Debugging and Performance Utilities" guide, remote devices must be in a "connected" state in order for coremltools to install the ModelRunner application. The devices in our system have a "paired" state, and I'm unable to set the them as "connected." The only way I know how to connect a device is to physically plug it in to a computer and open Xcode. I don't have physical access to the devices in the CI/CD system, and the host computer that interacts with them doesn't have Xcode installed. Here are some questions I've been looking into and would love some help answering: Has anyone managed to use the coremltools performance utilities in a similar system? Can you put a device in a "connected" state if you don't have physical access to the device and if you only have access to Xcode command line tools and not the Xcode app? Is it at all possible to install the coremltools ModelRunner application on a "paired" device, for example, by manually building the app and installing it with devicectl? Would other utilities, such as the MLModelBenchmarker work as expected if the app is installed this way? Thank you!