In 2025, on-device AI was a differentiator — something a few well-resourced teams added to stand out. In 2026, it is becoming the expected default for any mobile feature that involves understanding content, personalising experiences, or processing user input. The reasons are structural: Apple's Core ML 7 and Google's TensorFlow Lite 2.15 have made integration trivial, chip capabilities on even mid-range devices have caught up with the model sizes needed for useful inference, and privacy regulations in the EU, India, and the US are making cloud-based data processing harder to justify when a local alternative exists. This is what building with on-device AI actually looks like today.
Apple's Core ML 7, released with iOS 19 in early 2026, brings three advances that change what is practical on-device. Model compression via structured pruning and INT4 quantisation can reduce model size by 60-75% with minimal accuracy loss, making models that were previously too large for on-device use viable. On-device fine-tuning allows models to adapt to individual user behaviour without sending data to a server. And the Neural Engine on A18 and A18 Pro chips can sustain inference throughput sufficient for real-time video analysis.
Google's TensorFlow Lite 2.15 and the companion MediaPipe framework have followed a similar path. MediaPipe Tasks — the high-level SDK that wraps common ML tasks like object detection, pose estimation, hand tracking, and text classification — received a major update in Q1 2026. Tasks that previously required custom model integration and significant ML expertise are now two lines of code. The Snapdragon 8 Gen 4's Hexagon NPU and the Google Tensor G4 chip both deliver on-device inference fast enough for real-time applications.
The combined effect is that the hardware ceiling for on-device AI has risen above what most mobile applications actually need. The constraint in 2026 is no longer 'is this device fast enough?' but 'have we chosen the right model size and quantisation level for our use case?'
The GDPR enforcement actions of 2024-2025, combined with India's Digital Personal Data Protection Act and evolving US state privacy laws, have made cloud-based processing of sensitive user data significantly more legally complex. Any feature that sends audio, images, health data, or behavioural signals to a remote server requires clear consent, data retention policies, and in some jurisdictions a Data Processing Agreement.
On-device processing sidesteps most of this complexity. Data that never leaves the device does not trigger most of these requirements. For mobile apps in healthcare, finance, education, and any domain involving minors, this is not just a performance optimisation — it is a compliance strategy.
App store ratings have started reflecting this. Apps that are explicit about on-device processing in their privacy labels are seeing measurably better reviews and conversion rates from privacy-conscious users, particularly in European markets.
Document scanning and data extraction has become one of the most reliable on-device AI use cases. An app can capture an image of a receipt, business card, or form, extract structured data, and populate fields — all locally, in under 200ms, with no server round trip. Core ML's Vision framework handles document detection and text recognition; a small extraction model handles structure parsing.
Real-time content moderation for user-generated content is a second high-value use case. Rather than sending every image or text input to a cloud moderation API, a compact on-device classifier can do a first pass and only escalate uncertain cases to the server. This reduces API costs, improves latency, and handles the offline case gracefully.
Personalised recommendations without profiling are now feasible. A small collaborative filtering model trained on anonymised global data can be fine-tuned on-device using the individual user's behaviour — giving personalised results without any of that behaviour ever leaving the phone. This is precisely the privacy-preserving personalisation pattern that regulators are encouraging.
For iOS-primary teams, Core ML with the Vision and Natural Language frameworks handles the majority of use cases with the lowest integration overhead. Apple's Create ML tool can fine-tune models from your own data without Python or ML expertise. The ecosystem is tightly integrated with Xcode and the simulator, making iteration fast.
For cross-platform teams, TensorFlow Lite with MediaPipe Tasks gives the most consistent experience across Android and iOS. The Task API (image classification, object detection, text embedding, gesture recognition) covers the most common use cases. For anything custom, TensorFlow Lite's converter can quantise models from TensorFlow or PyTorch.
For teams that need to run LLM-class models on-device (summarisation, Q&A, code completion), the landscape in 2026 has a clear leader: Google's Gemini Nano (built into Pixel 9+ and selectively licensed to other OEMs) and Apple Intelligence models handle this on iOS. For cross-platform LLM inference without platform dependency, llama.cpp with 3B-7B parameter models is the most adopted open-source option.
Written by Evitras Team
Evitras Technologies · 2026-04-25
Ready to build something great?
Talk to the Evitras team about your next project.
Kotlin Multiplatform Mobile crossed from early-adopter to production-grade in 2026. Netflix, Spotify, and Cash App ship KMM codebases. Here is the honest guide to what KMP is, what it is not, and when to use it.
React Server Components are no longer experimental — they are the expected baseline. Here is what the server-first shift actually means for how teams build and ship React applications today.
The next UI paradigm is not a new layout trend — it is a shift from designing static interfaces to designing systems that adapt to user intent in real time. Here is what this looks like in production.
We publish new articles on technology, design, and strategy. Reach out to get notified when we publish something new.
We value your privacy
We use cookies to enhance your experience. Privacy Policy
