Apple just killed Core ML after 9 years. WWDC 2026 dropped Core AI — and it changes everything for iOS developers building with LLMs. Core ML was built in 2017 for image classifiers and tree ensembles running on the Neural Engine. It did that job well. But it was never designed for streaming token generation, agent-style tool calling, or large language models — which is exactly what every iOS dev is shipping now. Core AI gives you: - Native streaming token generation (no more hacks) - Structured JSON outputs from on-device models - Tool calling built into the inference pipeline - Session management for multi-turn agent interactions The kicker: the original 3-line Swift API hasn't changed. Dynamic Profiles let you swap tools, models, and instructions mid-session without tearing down the pipeline. If you have .mlmodel files in production, they'll keep working — but all new AI work starts on Core AI. Plan your migration now. Also new: Xcode 27 went fully agentic. The IDE can now take on entire coding tasks — not just suggest the next line. It ships with 20 MCP tools wired up, and any MCP-compatible agent (Claude Code, Codex) can invoke Xcode's build, test, and diagnostics ops. Tim Cook's last keynote. John Ternus takes over September 1. Apple is betting the farm on on-device AI. linkedin.com/pulse/wwdc-2026…

I know you work at Qualcomm, but that would be cool to have an MLModel for it 😁

day 2. I built my first KNN model, and it had a 93% accuracy score :) not bad for the first try in my opinion.😌 next up: Logistic regression model #buildinpublic #MachineLearning #mlmodel

EmotiBit Wearable biometric sensing for any project. emotibit.com/ EmotiBit is a wearable sensor module for capturing high-quality emotional, physiological, and movement data. Packed with sensors in a wearable form-factor, EmotiBit lets you immediately begin streaming 16 signals from the body. With EmotiBit you enjoy wireless data streaming to any platform or direct data recording to the built-in SD card. Customize the Arduino-compatible hardware and fully open-source software to meet any project needs. Currently EmotiBit supports 2 Adafruit Feathers, the Feather ESP32 Huzzah adafruit.com/product/3619 and the Feather M0 WiFi.adafruit.com/product/3044 The open connectivity of EmotiBit gives you flexibility in how you can access, analyze, and integrate biometric data. Out of the box you can connect to data streams using multiple protocols, including OSC, LSL, and UDP. You can also bring your data to life with Python and eight additional languages with the BrainFlow SDK. EmotiBit’s open-source ecosystem is constantly growing, and advanced users can extend connectivity options with firmware or software code changes to include protocols like TCP, MQTT, Bluetooth, LoRa and more! github.com/EmotiBit/EmotiBit… Validating EmotiBit, an open-source multi-modal sensor for capturing research-grade physiological signals from anywhere on the body sciencedirect.com/science/ar… Brainflow brainflow.org/ BrainFlow is a library intended to obtain, parse and analyze EEG, EMG, ECG, and other kinds of data from biosensors. It provides a uniform software development kit (SDK) to work with biosensors with a primary focus on neurointerfaces. Support for Python, C , Java, C#, Julia, Matlab, R, Typescript, and Rust. The same API across all bindings and uniform API for all devices. Switch devices without changes in code. github.com/brainflow-dev/bra… BrainFlow User API has three main modules: 🔸BoardShim to read data from a board, it calls methods from underlying BoardController library 🔸DataFilter to perform signal processing, it calls methods from underlying DataHandler library 🔸MLModel to calculate derivative metrics, it calls methods from underlying MLModule library These classes are independent, so if you want, you can use BrainFlow API only for data streaming and perform signal processing by yourself and vice versa. brainflow.readthedocs.io/en/… BrainFlow data acqusition API is board agnostic, so to select a specific board you need to pass BrainFlow’s board id to BoardShim’s constructor and an instance of BrainFlowInputParams structure which should hold information for your specific board. brainflow.readthedocs.io/en/… In BoardShim, all board data is returned as a 2d array. Rows in this array may contain timestamps, EEG and EMG data and so on. brainflow.readthedocs.io/en/… BrainFlow’s documentation brainflow.readthedocs.io/en/… EmotiBit board brainflow.readthedocs.io/en/… Integration with Game Engines Brainflow integrates with Unreal, Unity, and Cry game engines. brainflow.readthedocs.io/en/…

Core ML → Core AI isn't just a rename. Apple's building for LLMs and generative AI, not just classification. In code: MLModel becomes multi-modal. Your .mlmodel files need a new container format. Foundation Models gets first-class Xcode tooling. Abstract your ML layer now.

You are either training AI for free, Or leveraging paid opportunities. You can focus on opportunities that allow you to get paid as you train, Or join @ActionModelAI, a project that gives you fractional ownership. With Action Model, you'll also get paid for the tasks eventually, besides fractional ownership. Join Action mlModel today and become part of a robust AI training community

No, it’s using the ANE compiler but bypasses the MLModel CoreML API. The number of functions is big. Int8 is strange; you definitely need to quantize activations. I’ll compare vs. Apple example code later. (Busy finishing Qwen 3.5. Many Mamba conflicts with ANE limitations for tensors and groups.)

まずPythonで動くものを作ってからそれを元にiOSのSwiftのアプリをClaudeCodeで実装してもらうパターン多い。onnxのモデルの推論をCoreML（.mlmodel）に変更してもらったら爆速（3fps->30fps）になった

Check out my latest in @Forbes: When pitching the use of an #MLmodel, #datascientists rarely report on its potential value. They then experience an unnerving lack of confidence – followed by project failure. forbes.com/sites/ericsiegel/…

El inteligencia Artificial también engaña, el uso técnico que empleamos en cómputo es Machine Learning o MLModel

Here's the new frontier for #predictiveAI: a new kind of ML software for planning, selling, & greenlighting #MLmodel deployment. There's nothing like trying it out yourself. Participate in this popular hands-on @DataCamp webinar – now available on demand. datacamp.com/resources/webin…

🎉 I just finished Day 27 of the #100DaysOfSwiftUI at hackingwithswift.com/100/swi… via @twostraws --- Day 27 | .mlmodel 的使用 - 非静态属性初始化若依赖另一属性，则应将后者设为 static - Markdown 格式化字符串语法

Shouldn't a great #MLmodel be a sure bet to deploy? No. It won’t deploy because you haven't closed the sale. When you deliver a model, you haven't necessarily finished selling, no matter how much of a done deal it may seem to be. forbes.com/sites/ericsiegel/…

Learn how to transform Finance with AI in 3 Phases From environment readiness to AI integration with TensorFlow & Azure DevOps, and a cluster-based deployment using ADKAR, this journey redefines how finance embraces AI innovation. shorturl.at/owGtF #AIInFinance #MLModel

Synthetic data isn’t just smart, it’s ethical🙍‍♂️ It can simulate real-world risk scenarios to prepare AI models for success. 🏥 Healthcare: Train AI to detect sudden disease outbreaks. 🏛️ Government: Stimulate coordinated cyberattacks on public portals to improve response systems. 🏦 Banking/Insurance: Simulate rare transaction spikes to boost fraud detection accuracy. Want your AI to handle the unexpected? Start with smarter data. Start with #SyncoraAI #healthcareai #HIPAA #agenticAI #AIML #decentralizedAI  #syntheticdata #syntheticdatageneration #edgecase #mlmodel #financeai #governmentai #bankingai

Advanced client-side AI: Processing CoreML Models with the Vision Frameowork Setting up a Vision model This vision model setup is boilerplate common to any image processing work on Core ML, so this setup will be the same across MobileCLIP, MobileNet, YOLOv3, or DepthAnything. We fetch the bundled model and load it as a MLModel. Then, to enable it to work with the vision framework, we load this as a VNCoreMLModel. This is important because of the model inputs I mentioned before. The CLIP model only accepts inputs of a specific image resolution and colour space. The Vision framework automatically performs the heavy lifting of resizing, cropping, and colour space converting our image, so it matches the Core ML model’s input format. The embeddings, as pure JSON, are simpler. We just load the file into an array of TitleEmbeddings that store the names to use (e.g. “The Mighty X”), the keyword (mighty), and the embedding for this keyword - a large array of floats. Now that we have the models and embeddings loaded up on init, we can start processing images as they’re sent through. Processing an Image with Vision After selecting or taking a photo, we pass the UIImage to the model, along with a “name” parameter to personalise the result. To perform our image processing, we create a vision request using the model, then pass it a request handler made from the image. - VNCoreMLRequest wraps a Core ML model to return “observations” about images. - VNRequestHandler runs Vision observation requests on a single image and sends it through a callback - hence the withCheckedThrowingContinuation. Now that we’ve processed our image through the CoreML model, we can start to analyse the predictions in the results. Learn to implement client-side AI with CoreML and CLIP models in the full article: blog.jacobstechtavern.com/p/…

Decided to build a spam/ham classifier with an RNN for sequential text data. Hit some epoch errors .🙂.#MachineLearning #PyTorch #DataScience #buildinpublic #DeepLearning #AI #NeuralNetworks #MLModel

🎉 I just finished Day 26 of the #100DaysOfSwiftUI at hackingwithswift.com/100/swi… via @twostraws Today is All About - Date and DateComponents, And DatePicker Then we train our model using using Create ML Apple's built-in Application to Train a model and Export The Trained mlmodel

Totally forgot to clean outliers 🥲 Built a PyTorch model to predict house prices from area(sq ft), bedrooms, bathrooms, and stories.#MachineLearning #PyTorch #DataScience #buildinpublic #DeepLearning #AI #DataCleaning #Outliers #NeuralNetworks #MLModel #RealEstateAI