Apple researchers unlock faster AI on iPhones and Macs — NPUMoE engine lets Apple Silicon chips run advanced language models without draining battery

Apple researchers published a technical blueprint (NPUMoE) that lets the Neural Processing Unit (NPU) — a dedicated AI chip built into every Apple Silicon device — run Mixture-of-Experts language models (a type of AI that activates only the relevant processing pathways instead of using all of them). Previously, NPUs couldn't handle this workload because the routing logic created unpredictable shapes and unsupported operations.

The engine offloads dense computation to the NPU while keeping a CPU/GPU fallback, solving three bottlenecks: it handles the unpredictable expert routing that NPUs normally can't process, it avoids slow irregular operations (like top-k selection and data shuffling), and it reduces the overhead of launching thousands of tiny parallel tasks. Result: longer context windows and faster text generation happen entirely on the device's specialized AI chip instead of draining the main processor.

For users: AI features on your iPhone, MacBook Air, or iPad now respond faster and use less battery during long conversations or document analysis. For app makers: you can now deploy more capable AI assistants locally on-device without asking users to upgrade to a more expensive device model — opening advanced AI features to the full Apple Silicon install base rather than just high-end machines.