MediaTek 2nm Chip Runs a 30B AI Model on Your Phone
2 min readThe phone in your pocket is about to run a 30 billion parameter language model without calling a data center. Taiwan’s MediaTek has unveiled the Dimensity 9600 Pro, the first smartphone processor built on a 2nm manufacturing process, and its headline feature is not raw speed. It is how much AI the chip can hold locally.
What Changed in On-Device AI
Until recently, serious AI on a phone meant a round trip to a server. The model lived in a data center, your request travelled there, and the answer came back. That works, but it costs the vendor money on every query, it stops working offline, and it means personal data leaves the device.
Phone makers have been chipping away at that by running small models locally, typically in the 3 to 8 billion parameter range. Useful for summarizing a message. Not enough for the multi-step agent behavior the industry keeps promising.
What MediaTek Announced
The Dimensity 9600 Pro is fabricated on TSMC’s 2nm N2P process and pairs two neural processing units with what MediaTek calls an Agentic AI Engine, a block designed for models that perceive, reason and then act rather than simply answer. The company says the chip supports on-device models up to 30 billion parameters.
The efficiency numbers matter as much as the ceiling. MediaTek claims its second-generation Super Efficient NPU cuts power draw by 40 percent for always-on AI tasks, while the NPU 1090 delivers 51 percent higher prefill performance and 55 percent more tokens generated per watt. CPU gains are smaller, roughly 17 percent single-core, but multi-core power consumption drops 61 percent. A 3nm sibling, the Dimensity 9600M, carries the same Agentic AI Engine down to cheaper handsets. Commercial devices ship this quarter.
Why It Matters
If a 30 billion parameter model genuinely runs on a handset, the economics of consumer AI shift. Inference that costs a cloud provider real money per query becomes free once the phone is sold, and an assistant that works in airplane mode with no data leaving the device is a materially different product from one that does not.
It also puts Taiwan at both ends of the chain. TSMC makes the 2nm wafers, MediaTek designs the chip, and together they set the floor for what Android flagships can do in 2027. Watch whether the first shipping phones actually load models near that 30 billion ceiling, or settle for something far smaller.