Hi all, I've been working on a pure-Swift port of Google's Gemma 4 text decoder that plugs into mlx-swift-lm as a sidecar model registration. Sharing it here in case anyone else hit the same wall I did, and to get feedback from the MLX team and the community before I propose anything upstream. Repo: https://github.com/yejingyang8963-byte/Swift-gemma4-core Why As of mlx-swift-lm 2.31.x, Gemma 4 isn't supported out of the box. The obvious workaround — reusing the Gemma 3 text implementation with a patched config — fails at weight load because Gemma 4 differs from Gemma 3 in several structural places. The chat-template path through swift-jinja 1.x also silently corrupts the prompt, so the model loads but generates incoherent text. What's in the package A from-scratch Swift implementation of the Gemma 4 decoder (Configuration, Layers, Attention, MLP, RoPE, DecoderLayer) Per-Layer Embedding (PLE) support — the shared embedding table that feeds every decoder layer through a gated MLP as a third residual KV sharing across the back half of the decoder, threaded through the forward pass via a donor table with a single global rope offset A custom Gemma4ProportionalRoPE class for the partial-rotation rope type that initializeRope doesn't currently recognize A chat-template bypass that builds the prompt as a literal string with the correct turn markers and encodes via tokenizer.encode(text:), matching Python mlx-lm's apply_chat_template byte-for-byte Measured on iPhone (A-series, 7.4 GB RAM) Model: mlx-community/gemma-4-e2b-it-4bit Warm load: ~6 s Memory after load: 341–392 MB Time to first token (end-to-end, 333-token system prompt): 2.82 s Generation throughput: 12–14 tok/s What I'd love feedback on Is the sidecar registration pattern the right way to extend mlx-swift-lm with new model families, or is there a more idiomatic path I missed? The chat-template bypass works but feels like a workaround. Is the right long-term fix in swift-jinja, in the tokenizer, or somewhere else entirely? Anyone running into the same PLE / KV-sharing issues on other Gemma-family checkpoints? I'd like to make sure the implementation generalizes beyond E2B before tagging a 0.2.0. Happy to open a PR against mlx-swift-lm if the maintainers think any of this belongs upstream. Thanks for reading.

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.

This YouTube video is very interesting, discussing Swift's power and its potential to replace Python. Here is the link. https://youtu.be/6ZGlseSqar0?si=pzZVq9FKsveca4kA

Things I did: created an Intents Extension target added "Supported Intents" to both my main app target and the intent extension, with "INAddTasksIntent" and "INCreateNoteIntent" created the AppIntentVocabulary in my main app target created the handlers in the code in the Intents Extension target class AddTaskIntentHandler: INExtension, INAddTasksIntentHandling { func resolveTaskTitles(for intent: INAddTasksIntent) async -> [INSpeakableStringResolutionResult] { if let taskTitles = intent.taskTitles { return taskTitles.map { INSpeakableStringResolutionResult.success(with: $0) } } else { return [INSpeakableStringResolutionResult.needsValue()] } } func handle(intent: INAddTasksIntent) async -> INAddTasksIntentResponse { // my code to handle this... let response = INAddTasksIntentResponse(code: .success, userActivity: nil) response.addedTasks = tasksCreated.map { INTask( title: INSpeakableString(spokenPhrase: $0.name), status: .notCompleted, taskType: .completable, spatialEventTrigger: nil, temporalEventTrigger: intent.temporalEventTrigger, createdDateComponents: DateHelper.localCalendar().dateComponents([.year, .month, .day, .minute, .hour], from: Date.now), modifiedDateComponents: nil, identifier: $0.id ) } return response } } class AddItemIntentHandler: INExtension, INCreateNoteIntentHandling { func resolveTitle(for intent: INCreateNoteIntent) async -> INSpeakableStringResolutionResult { if let title = intent.title { return INSpeakableStringResolutionResult.success(with: title) } else { return INSpeakableStringResolutionResult.needsValue() } } func resolveGroupName(for intent: INCreateNoteIntent) async -> INSpeakableStringResolutionResult { if let groupName = intent.groupName { return INSpeakableStringResolutionResult.success(with: groupName) } else { return INSpeakableStringResolutionResult.needsValue() } } func handle(intent: INCreateNoteIntent) async -> INCreateNoteIntentResponse { do { // my code for handling this... let response = INCreateNoteIntentResponse(code: .success, userActivity: nil) response.createdNote = INNote( title: INSpeakableString(spokenPhrase: itemName), contents: itemNote.map { [INTextNoteContent(text: $0)] } ?? [], groupName: INSpeakableString(spokenPhrase: list.name), createdDateComponents: DateHelper.localCalendar().dateComponents([.day, .month, .year, .hour, .minute], from: Date.now), modifiedDateComponents: nil, identifier: newItem.id ) return response } catch { return INCreateNoteIntentResponse(code: .failure, userActivity: nil) } } } uninstalled my app restarted my physical device and simulator Yet, when I say "Remind me to buy dog food in Index" (Index is the name of my app), as stated in the examples of INAddTasksIntent, Siri proceeds to say that a list named "Index" doesn't exist in apple Reminders app, instead of processing the request in my app. Am I missing something?

I have a series of shortcuts that I’ve written that use the “Use Model” action to do various things. For example, I have a shortcut “Clipboard Markdown to Notes” that takes the content of the clipboard, creates a new note in Notes, converts the markdown content to rich text, adds it to the note etc. One key step is to analyze the markdown content with “Use Model” and generate a short descriptive title for the note. I use the on-device model for this, but sometimes the content and prompt exceed the context window size and the action fails with an error message to that effect. In that case, I’d like to either repeat the action using the Cloud model, or, if the error was a refusal, to prompt the user to enter a title to use. I‘ve tried using an IF based on whether the response had any text in it, but that didn’t work. No matter what I’ve tried, I can’t seem to find a way to catch the error from Use Model, determine what the error was, and take appropriate action. Is there a way to do this? (And by the way, a huge ”thank you” to whoever had the idea of making AppIntents visible in Shortcuts and adding the Use Model action — has made a huge difference already, and it lets us see what Siri will be able to use as well.)

Hi, I'm trying to add FoundationModels to an older project but always get the following error: "Unable to resolve 'dependency' 'FoundationModels' import FoundationModels" The error comes and goes while its compiling and then doesn't run the app. I have my target set to 26.0 (and can't go any higher) and am using Xcode 26 (17E192). Is anyone else having this issue? Thanks, Dan Uff

Hello! After the 26.4 update I get a huge number of LanguageModelSession.GenerationError.refusal errors when using guided generation Generables for inexplicable reasons. Such errors also occur, if I want to cast a response to boolean by using 'generating: Bool.self'. The explanation generated on the grounds of the error always looks like this: Response(userPrompt: "", duration: 0.230917542, promptTokenCount: Optional(66), responseTokenCount: Optional(11), feedbackAttachment: nil, content: "I apologize, but I cannot fulfill this request.", rawContent: "I apologize, but I cannot fulfill this request.", transcriptEntries: ArraySlice([])) All the prompts and Generables I use are definitely not profane. Before 26.4 such errors on the same prompts and Generables never occurred. The 26.4 update rendered those features unusable to me. Is this a known bug or what am I doing wrong?

After installing iOS 26.4 the Foundation Models instruction following and tool calling capabilities have been degraded significantly. The model is not usable anymore. Examples: This works: "Is the car plugged in?" This does not work: "Tell me if the car is plugged in" Anything with the work "frunk" (front trunk) triggers Guardrail Violation. Phrases like "Lock Pride" also trigger Guardrail Violation (Pride is the name of the car). Tool calling only works half the time for really obvious things.

When I am doing an uncached load of CoreML model on ANE, I received this warning in Xcode console Type of hiddenStates in function main's I/O contains unknown strides. Using unknown strides for MIL tensor buffers with unknown shapes is not recommended in E5ML. Please use row_alignment_in_bytes property instead. Refer to https://e5-ml.apple.com/more-info/memory-layouts.html for more information. However, the web link does not seem to be working. Where can I find more information about about this and how can I fix it?

I am a professional user of the 2019 Mac Pro (7,1) with dual AMD Radeon Pro W6900X MPX modules (32GB VRAM each). This hardware is designed for high-performance compute, but it is currently crippled for modern local LLM/AI workloads under Linux due to Apple's EFI/PCIe routing restrictions. Core Issue: rocminfo reports "No HIP GPUs available" when attempting to use ROCm/amdgpu on Linux Apple's custom EFI firmware blocks full initialization of professional GPU compute assets The dual W6900X GPUs have 64GB combined VRAM and high-bandwidth Infinity Fabric Link, but cannot be fully utilized for local AI inference/training My Specific Request: Apple should provide an official, one-click deployable application that enables full utilization of dual W6900X GPUs for local large language model (LLM) inference and training under Linux. This application must: Fully initialize both W6900X GPUs via HIP/ROCm, establishing valid compute contexts Bypass artificial EFI/PCIe routing restrictions that block access to professional GPU resources Provide a stable, user-friendly one-click deployment experience (similar to NVIDIA's AI Enterprise or AMD's ROCm Hub) Why This Matters: The 2019 Mac Pro is Apple's flagship professional workstation, marketed for compute-intensive workloads. Its high-cost W6900X GPUs should not be locked down for modern AI/LLM use cases. An official one-click deployment solution would demonstrate Apple's commitment to professional AI and unlock significant value for professional users. I look forward to Apple's response and a clear roadmap for enabling this critical capability. #MacPro #Linux #ROCm #LocalLLM #W6900X #CoreML

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.

Following WWDC24 video "Discover Swift enhancements in the Vision framework" recommendations (cfr video at 10'41"), I used the following code to perform multiple new iOS 18 `RecognizedTextRequest' in parallel. Problem: if more than 2 request are run in parallel, the request will hang, leaving the app in a state where no more requests can be started. -> deadlock I tried other ways to run the requests, but no matter the method employed, or what device I use: no more than 2 requests can ever be run in parallel. func triggerDeadlock() {} try await withThrowingTaskGroup(of: Void.self) { group in // See: WWDC 2024 Discover Siwft enhancements in the Vision framework at 10:41 // ############## THIS IS KEY let maxOCRTasks = 5 // On a real-device, if more than 2 RecognizeTextRequest are launched in parallel using tasks, the request hangs // ############## THIS IS KEY for idx in 0..<maxOCRTasks { let url = ... // URL to some image group.addTask { // Perform OCR let _ = await performOCRRequest(on: url: url) } } var nextIndex = maxOCRTasks for try await _ in group { // Wait for the result of the next child task that finished if nextIndex < pageCount { group.addTask { let url = ... // URL to some image // Perform OCR let _ = await performOCRRequest(on: url: url) } nextIndex += 1 } } } } // MARK: - ASYNC/AWAIT version with iOS 18 @available(iOS 18, *) func performOCRRequest(on url: URL) async throws -> [RecognizedText] { // Create request var request = RecognizeTextRequest() // Single request: no need for ImageRequestHandler // Configure request request.recognitionLevel = .accurate request.automaticallyDetectsLanguage = true request.usesLanguageCorrection = true request.minimumTextHeightFraction = 0.016 // Perform request let textObservations: [RecognizedTextObservation] = try await request.perform(on: url) // Convert [RecognizedTextObservation] to [RecognizedText] return textObservations.compactMap { observation in observation.topCandidates(1).first } } I also found this Swift forums post mentioning something very similar. I also opened a feedback: FB17240843

Hi, I’m working on a real-time head/face tracking pipeline using a standard 2D RGB camera, and I’m trying to better understand how ARKit achieves such stable and responsive results in comparable conditions. To clarify upfront: I’m specifically interested in RGB-only tracking and the underlying vision/ML pipeline. I’m not using TrueDepth or any depth/IR-based sensors, and I’d like to understand how similar stability and responsiveness can be achieved under those constraints. In my current setup, I estimate head pose from RGB frames (facial landmarks + PnP) and apply temporal filtering (e.g., exponential smoothing and Kalman filtering). This significantly reduces jitter, but introduces noticeable latency, especially during faster head movements. What stands out in ARKit is that it appears to maintain both: Very low jitter Very low perceived latency even when operating with camera input alone. I’m trying to understand what techniques might contribute to this behavior. In particular: Does ARKit use predictive tracking (e.g., velocity or acceleration-based pose extrapolation) to compensate for camera and processing delays in RGB-only scenarios? Are there recommended strategies for balancing temporal smoothing and responsiveness without introducing visible lag in camera-based pose estimation pipelines? Is the tracking pipeline internally decoupled from rendering (e.g., asynchronous processing with prediction applied at render time)? Are there general best practices for minimizing end-to-end latency in vision-based head tracking systems beyond standard filtering approaches? I understand that implementation details may not be public, but any high-level insights or pointers would be greatly appreciated. Thanks!

Hello, I'm using videotoolbox superresolution API in MACOS 26: https://developer.apple.com/documentation/videotoolbox/vtsuperresolutionscalerconfiguration/downloadconfigurationmodel(completionhandler:)?language=objc, when using swift, it's ok, when using objective-c, I get error when downloading model with downloadConfigurationModelWithCompletionHandler: [Auto] MA-auto{_failedLockContent} | failure reported by server | error:[com.apple.MobileAssetError.AutoAsset:MissingReference(6111)] [Auto] MA-auto{_failedLockContent} | failure reported by server | error:[com.apple.MobileAssetError.AutoAsset:UnderlyingError(6107)_1_com.apple.MobileAssetError.Download:47] Download completion handler called with error: The operation couldnxe2x80x99t be completed. (VTFrameProcessorErrorDomain error -19743.)

Hello, Could you please provide details for maximum string length of the prompt and the title when using ImageCreator and the method extracted(from:title:)? static func extracted( from text: String, title: String? = nil ) -> ImagePlaygroundConcept Any additional details or example of prompt and title would help. Additionally, are ImagePlaygroundStyle.animation, ImagePlaygroundStyle.illustration and ImagePlaygroundStyle.sketch all available when using extracted(from:title:)? I am trying to generate images programmatically and would appreciate your guidance. Thank you.

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!

While analyzing system power on an M4 Max under GPU-heavy compute workloads, I noticed that the the GPU power reported by powermetrics does not come anywhere close to total system DC power reported by the SMC counter PDTR (as used by utilities like mactop). For example, in a heavy GPU workload, powermetrics would report a 65W idle-load delta on the GPU, but at the same time system DC power would rise by 179W, leaving 114W or nearly 2/3 of total system DC power on a Mac Studio M4 Max unexplained. From measurements, the difference appears to correlate with the amount of on-chip data movement (for example, varying bytes-per-FLOP in the workload changes the observed gap). Using SMC and IOReport, I was able to reverse engineer an energy model for the GPU that explains almost all of the energy flow with less than 2% error on the workload I studied. The result is a simple two-term energy roofline model: P_GPU (GPU_combined term in the plot) ≈ a * bytes + b * FLOPs with: ~5 pJ/byte for SRAM movement ~2.7 pJ/FLOP for compute. Has anyone observed similar behavior, or is there guidance on how GPU power reported by IOReport/powermetrics should be interpreted relative to total system power? In particular, I’m interested in whether certain classes of GPU activity may not be attributed to the GPU component in IOReport. Full details with the methodology and results are available here: https://youtu.be/HKxIGgyeISM

We built an open-source macOS menu bar app that turns speech into text and pastes it into the active app — using SpeechAnalyzer for on-device transcription, ScreenCaptureKit + Vision for screen-aware context, and FluidAudio for speaker diarization in meeting mode. Here's what we learned shipping it on macOS 26. GitHub: github.com/Marvinngg/ambient-voice Architecture The app has two modes: hotkey dictation (press to talk, release to inject) and meeting recording (continuous transcription with a floating panel). Dictation Mode Audio capture uses AVCaptureSession (more on why below). The captured audio feeds into SpeechAnalyzer via an AsyncStream: let transcriber = SpeechTranscriber( locale: locale, transcriptionOptions: [], reportingOptions: [.volatileResults, .alternativeTranscriptions], attributeOptions: [.audioTimeRange, .transcriptionConfidence] ) let analyzer = SpeechAnalyzer(modules: [transcriber]) let (inputSequence, inputBuilder) = AsyncStream.makeStream() try await analyzer.start(inputSequence: inputSequence) While recording, we capture a screenshot of the focused window using ScreenCaptureKit, run Vision OCR (VNRecognizeTextRequest), extract keywords, and inject them into SpeechAnalyzer as contextual bias: let context = AnalysisContext() context.contextualStrings[.general] = ocrKeywords try await analyzer.setContext(context) This improves accuracy for technical terms and proper nouns visible on screen. If your screen shows "SpeechAnalyzer", saying it out loud is more likely to be transcribed correctly. After transcription, an optional L2 step sends the text through a local LLM (ollama) for spoken-to-written cleanup, then CGEvent simulates Cmd+V to paste into the active app. Meeting Mode Meeting mode forks the same audio stream to two consumers: SpeechAnalyzer — real-time streaming transcription, displayed in a floating NSPanel FluidAudio buffer — accumulates 16kHz Float32 mono samples for batch speaker diarization after recording stops When the user ends the meeting, FluidAudio's performCompleteDiarization() runs on the accumulated audio. We align transcription segments with speaker segments using audioTimeRange overlap matching — each transcription segment gets assigned the speaker ID with the most time overlap. Results export to Markdown. Pitfalls We Hit on macOS 26 1. AVAudioEngine installTap doesn't fire with Bluetooth devices We started with AVAudioEngine.inputNode.installTap() for audio capture. It worked fine with built-in mics but the tap callback never fired with Bluetooth devices (tested with vivo TWS 4 Hi-Fi). Fix: switched to AVCaptureSession. The delegate callback captureOutput(_:didOutput:from:) fires reliably regardless of audio device. The tradeoff is you get CMSampleBuffer instead of AVAudioPCMBuffer, so you need a conversion step. 2. NSEvent addGlobalMonitorForEvents crashes Our global hotkey listener used NSEvent.addGlobalMonitorForEvents. On macOS 26, this crashes with a Bus error inside GlobalObserverHandler — appears to be a Swift actor runtime issue. Fix: switched to CGEventTap. Works reliably, but the callback runs on a CFRunLoop context, which Swift doesn't recognize as MainActor. 3. CGEventTap callbacks aren't on MainActor If your CGEventTap callback touches any @MainActor state, you'll get concurrency violations. The callback runs on whatever thread owns the CFRunLoop. Fix: bridge with DispatchQueue.main.async {} inside the tap callback before touching any MainActor state. 4. CGPreflightScreenCaptureAccess doesn't request permission We used CGPreflightScreenCaptureAccess() as a guard before calling ScreenCaptureKit. If it returned false, we'd bail out. The problem: this function only checks — it never triggers macOS to add your app to the Screen Recording permission list. Chicken-and-egg: you can't get permission because you never ask for it. Fix: call CGRequestScreenCaptureAccess() at app startup. This adds your app to System Settings → Screen Recording. Then let ScreenCaptureKit calls proceed without the preflight guard — SCShareableContent will also trigger the permission prompt on first use. 5. Ad-hoc signing breaks TCC permissions on every rebuild During development, codesign --sign - (ad-hoc) generates a different code directory hash on every build. macOS TCC tracks permissions by this hash, so every rebuild = new app identity = all permissions reset. Fix: sign with a stable certificate. If you have an Apple Development certificate, use that. The TeamIdentifier stays constant across rebuilds, so TCC permissions persist. We also discovered that launching via open WE.app (LaunchServices) instead of directly executing the binary is required — otherwise macOS attributes TCC permissions to Terminal, not your app. Benchmarks We ran end-to-end benchmarks on public datasets (Mac Mini M4 16GB, macOS 26): Transcription (SpeechAnalyzer, AliMeeting Chinese): • Near-field CER 34% (excluding outliers ~25%) • Far-field CER 40% (single channel, no beamforming, >30% overlap) • Processing speed 74-89x real-time Speaker diarization (FluidAudio offline): • AMI English 16 meetings: avg DER 23.2% (collar=0.25s, ignoreOverlap=True) • AliMeeting Chinese 8 meetings: DER 48.5% (including overlap regions) • Memory: RSS ~500MB, peak 730-930MB Full evaluation methodology, scripts, and raw results are in the repo. Open Source The project is MIT licensed: github.com/Marvinngg/ambient-voice It includes the macOS client (Swift 6.2, SPM), server-side distillation/training scripts (Python), and a complete evaluation framework with reproducible benchmarks. Feedback and contributions welcome.

We built an open-source macOS menu bar app that turns speech into text and pastes it into the active app — using SpeechAnalyzer for on-device transcription, ScreenCaptureKit + Vision for screen-aware context, and FluidAudio for speaker diarization in meeting mode. Here's what we learned shipping it on macOS 26. GitHub: github.com/Marvinngg/ambient-voice Architecture The app has two modes: hotkey dictation (press to talk, release to inject) and meeting recording (continuous transcription with a floating panel). Dictation Mode Audio capture uses AVCaptureSession (more on why below). The captured audio feeds into SpeechAnalyzer via an AsyncStream: let transcriber = SpeechTranscriber( locale: locale, transcriptionOptions: [], reportingOptions: [.volatileResults, .alternativeTranscriptions], attributeOptions: [.audioTimeRange, .transcriptionConfidence] ) let analyzer = SpeechAnalyzer(modules: [transcriber]) let (inputSequence, inputBuilder) = AsyncStream.makeStream() try await analyzer.start(inputSequence: inputSequence) While recording, we capture a screenshot of the focused window using ScreenCaptureKit, run Vision OCR (VNRecognizeTextRequest), extract keywords, and inject them into SpeechAnalyzer as contextual bias: let context = AnalysisContext() context.contextualStrings[.general] = ocrKeywords try await analyzer.setContext(context) This improves accuracy for technical terms and proper nouns visible on screen. If your screen shows "SpeechAnalyzer", saying it out loud is more likely to be transcribed correctly. After transcription, an optional L2 step sends the text through a local LLM (ollama) for spoken-to-written cleanup, then CGEvent simulates Cmd+V to paste into the active app. Meeting Mode Meeting mode forks the same audio stream to two consumers: SpeechAnalyzer — real-time streaming transcription, displayed in a floating NSPanel FluidAudio buffer — accumulates 16kHz Float32 mono samples for batch speaker diarization after recording stops When the user ends the meeting, FluidAudio's performCompleteDiarization() runs on the accumulated audio. We align transcription segments with speaker segments using audioTimeRange overlap matching — each transcription segment gets assigned the speaker ID with the most time overlap. Results export to Markdown. Pitfalls We Hit on macOS 26 1. AVAudioEngine installTap doesn't fire with Bluetooth devices We started with AVAudioEngine.inputNode.installTap() for audio capture. It worked fine with built-in mics but the tap callback never fired with Bluetooth devices (tested with vivo TWS 4 Hi-Fi). Fix: switched to AVCaptureSession. The delegate callback captureOutput(_:didOutput:from:) fires reliably regardless of audio device. The tradeoff is you get CMSampleBuffer instead of AVAudioPCMBuffer, so you need a conversion step. 2. NSEvent addGlobalMonitorForEvents crashes Our global hotkey listener used NSEvent.addGlobalMonitorForEvents. On macOS 26, this crashes with a Bus error inside GlobalObserverHandler — appears to be a Swift actor runtime issue. Fix: switched to CGEventTap. Works reliably, but the callback runs on a CFRunLoop context, which Swift doesn't recognize as MainActor. 3. CGEventTap callbacks aren't on MainActor If your CGEventTap callback touches any @MainActor state, you'll get concurrency violations. The callback runs on whatever thread owns the CFRunLoop. Fix: bridge with DispatchQueue.main.async {} inside the tap callback before touching any MainActor state. 4. CGPreflightScreenCaptureAccess doesn't request permission We used CGPreflightScreenCaptureAccess() as a guard before calling ScreenCaptureKit. If it returned false, we'd bail out. The problem: this function only checks — it never triggers macOS to add your app to the Screen Recording permission list. Chicken-and-egg: you can't get permission because you never ask for it. Fix: call CGRequestScreenCaptureAccess() at app startup. This adds your app to System Settings → Screen Recording. Then let ScreenCaptureKit calls proceed without the preflight guard — SCShareableContent will also trigger the permission prompt on first use. 5. Ad-hoc signing breaks TCC permissions on every rebuild During development, codesign --sign - (ad-hoc) generates a different code directory hash on every build. macOS TCC tracks permissions by this hash, so every rebuild = new app identity = all permissions reset. Fix: sign with a stable certificate. If you have an Apple Development certificate, use that. The TeamIdentifier stays constant across rebuilds, so TCC permissions persist. We also discovered that launching via open WE.app (LaunchServices) instead of directly executing the binary is required — otherwise macOS attributes TCC permissions to Terminal, not your app. Benchmarks We ran end-to-end benchmarks on public datasets (Mac Mini M4 16GB, macOS 26): Transcription (SpeechAnalyzer, AliMeeting Chinese): • Near-field CER 34% (excluding outliers ~25%) • Far-field CER 40% (single channel, no beamforming, >30% overlap) • Processing speed 74-89x real-time Speaker diarization (FluidAudio offline): • AMI English 16 meetings: avg DER 23.2% (collar=0.25s, ignoreOverlap=True) • AliMeeting Chinese 8 meetings: DER 48.5% (including overlap regions) • Memory: RSS ~500MB, peak 730-930MB Full evaluation methodology, scripts, and raw results are in the repo. Open Source The project is MIT licensed: github.com/Marvinngg/ambient-voice It includes the macOS client (Swift 6.2, SPM), server-side distillation/training scripts (Python), and a complete evaluation framework with reproducible benchmarks. Feedback and contributions welcome.

In macOS Tahoe 26.2 an RDMA capability was added for Thunderbolt-5 interfaces. This has been demonstrated to significantly decrease the latency and maintain bandwidth for "clustered" Apple Silicon devices with TB5. What is the ideal and the maximum RDMA burst width for transfers over RDMA-enabled Thunderbolt-5 interfaces?