MediaTek boosts on-device AI speed by 50% using Phi models from Microsoft

MediaTek chipset enabling on-device AI features powered by Phi

MediaTek enhanced chipset performance for on-device AI using Microsoft's Phi models, enabling real-time functionality without internet connectivity. The deployment achieved 50% faster processing, 80% improved prompt performance, and 30% greater power efficiency.

50%
Faster on-device AI processing
80%
Improved prompt performance
30%
Greater power efficiency

MediaTek is a leading semiconductor company designing system-on-chip solutions for mobile devices and IoT applications, facing rising customer demand for advanced AI capabilities that run directly on handheld devices without internet connectivity. Customers wanted real-time features like translation and multimodal assistance with minimal latency and high accuracy — an experience MediaTek needed to deliver to maintain its competitive edge in the smartphone market.

MediaTek selected Microsoft Azure AI Foundry and integrated Phi models from the Azure AI Model Catalog into its Dimensity 9400 chipsets. The team leveraged Phi-3.5 mini for rapid customization and applied speculative decoding to optimize token processing speed. Product Manager Yannic Peng noted: “Our customers use the Phi-3.5 mini model, which allows them to rapidly customize their product.” Edge AI optimization kept sensitive data processing local, reducing latency and eliminating constant cloud dependency.

The deployment delivered 50% faster on-device AI processing (18 tokens per second versus the previous 12), 80% improved language model prompt performance, and 30% greater power efficiency. MediaTek plans to expand capabilities this year by adding the Phi-4-multimodal model to enable speech and vision inputs, further enhancing the user experience on consumer devices.

Try it yourself

Build with the same model powering this story.

Try Phi

Ready to write your success story?

Connect with our team to explore how Foundry Labs innovations can accelerate your business.

Join our community →