Qualcomm's Snapdragon 8 Elite Gen 6 chips target on-device agents, with a 30B model on the Extreme version

TechCrunch's Ivan Mehta reports that at its Snapdragon Summit, Qualcomm announced two flagship smartphone processors, the Snapdragon 8 Elite Gen 6 and the Snapdragon 8 Elite Extreme Gen 6. Both are framed around better AI features and personalization for AI agents, including local scribe features, speaker identification and a sensing hub.

The Extreme version can run a 30-billion-parameter mixture-of-experts model on the device. Motorola's Signature 27 is an early phone using the Extreme Gen 6.

Qualcomm's bet is that most people will meet AI agents on phones first rather than on dedicated gadgets. Hardware that can run useful models locally lets agents work without sending every request to the cloud, which can help with privacy, latency and cost.

Dedicated devices are still arriving. Meta's Muse Charm includes its own 5G modem and screen, and Microsoft has dropped the Copilot+ PC label from new Surfaces powered by Qualcomm's Snapdragon X2 Plus, even though their NPUs exceed the spec. Qualcomm's phone chips aim to make AI capability standard rather than a premium label. TechCrunch's comparison point is Apple, which runs its own on-device mixture-of-experts stack; how the Extreme's 30-billion-parameter model holds up against Apple's approach in daily use is one of the open questions.

On-device processing is a counterweight to cloud-based agents. Google has taken a hybrid approach with encrypted server-side memory in Private AI Compute, while Nvidia is promoting local AI with the open-source community. Speaker identification on phones also parallels Nvidia's open Nemotron 3 Diarization model.

Local speaker identification and transcription raise the same consent questions as other always-listening devices, as covered in everything is spying on you. The Motorola Signature 27's ship date and independent battery and thermal tests with local agents running, which phone makers expose sensing-hub features to developers, how the 30-billion-parameter model performs in practice and whether new features are defaults or buried settings are what to watch. More in our AI coverage and tools coverage.