MediaTek and Microsoft unveil a breakthrough in Android AI development with the Phi-4-mini models optimized for MediaTek’s NPUs. The NeuroPilot SDK enables fast, privacy-focused AI apps running offline on devices like smartphones and IoT, transforming on-device intelligence and developer workflows. Unique :

Transforming Android Development with MediaTek & Microsoft’s Phi Models
Imagine running advanced AI apps like intelligent copilots and Retrieval-Augmented Generation (RAG) offline on Android devices. Thanks to the rapid evolution of Neural Processing Units (NPUs), this is now a reality. MediaTek and Microsoft have teamed up to bring powerful AI directly to your phone, tablet, or IoT device—no cloud required.
What’s New: Phi-4-mini Optimized for MediaTek NPUs
MediaTek’s latest Dimensity 9400 and 9400+ platforms, combined with the Dimensity GenAI Toolkit 0, now support Microsoft’s Phi-4-mini and Phi-4-mini-reasoning models. These models are finely tuned for MediaTek’s NPUs, enabling lightning-fast AI processing on-device.
“Prefill speed is over 800 tokens per second, and decode speed exceeds 21 tokens per second.”
This means developers can build AI experiences that are not only fast but also privacy-preserving, since everything runs locally without cloud dependency.
Major Updates: MediaTek NeuroPilot SDK and Toolchain
The NeuroPilot SDK is a game-changer for AI developers. It offers a full suite of tools to convert, quantize, compile, and deploy AI models efficiently across all MediaTek AI-capable hardware.
Supporting a “code once, deploy everywhere” approach, the SDK works seamlessly on smartphones, tablets, automotive systems, smart home devices, and IoT gadgets. Integration with Android and Linux ecosystems ensures broad compatibility and optimized performance.
“Developers don’t need specialized hardware expertise to rapidly prototype and deploy custom AI solutions.”
Seamless AI Deployment on Edge Devices
With demos showcasing Phi-4-mini-reasoning and Phi-4-mini models, developers can now bring advanced reasoning and instruction-following AI directly to edge devices. This opens doors to smarter assistants, productivity tools, and context-aware automation that work offline.
Real-World Use Cases: Intelligent On-Device Assistants
Imagine an AI assistant that understands your personal documents, PDFs, or notes without sending data to the cloud. With on-device RAG powered by Phi-4-mini, apps can deliver personalized, private, and lightning-fast information retrieval.
This technology enables personalized assistants, offline knowledge hubs, and smart summarization tools that keep your data secure and accessible anytime.
Stay Ahead: Upcoming Events and Resources
For developers eager to dive deeper, Microsoft Build 2025 and Computex Taipei offer key sessions on Azure AI Foundry and MediaTek’s AI hardware innovations. Don’t miss out on hands-on labs and expert insights to master building next-gen AI apps.
Explore the Phi-4 model family on Azure AI Foundry and HuggingFace, and get your hands on the Phi Cookbook for practical guides and code repositories.
In short, MediaTek and Microsoft are pushing the boundaries of AI on Android devices. Expect faster, smarter, and more private AI experiences—right in your pocket.
From the New blog articles in Microsoft Community Hub