Qualcomm’s new flagship chips push more AI processing onto phones

Introduction

Qualcomm has announced two flagship mobile processors at its annual Snapdragon Summit: the Snapdragon 8 Elite Gen 6 and the Snapdragon 8 Elite Extreme Gen 6. The announcement is less about headline CPU performance than about moving more AI functions directly onto smartphones. Qualcomm is targeting personalized agents, voice interaction, device-side memory, and more sophisticated camera processing.

Key takeaways

• A new sensing hub: The chips can run small models of up to roughly 200 million parameters locally. Qualcomm says this can support a personal scribe, speaker differentiation, and a memory of usage patterns that can improve future suggestions and task automation.
• End-to-end voice agents: The company says the new silicon can support a complete voice-in, voice-out agent, potentially allowing users to interact with a phone in a more natural way without treating speech recognition and response generation as separate features.
• More efficient local inference: The Snapdragon 8 Elite Gen 6 includes a new accelerator element designed to run models more efficiently. The Extreme version can run a 30-billion-parameter mixture-of-experts model on-device. Although the model has 30B parameters overall, only a subset is activated for a given task, reducing the amount of computation required compared with activating the entire model.
• Camera and video upgrades: The new CPU offers pixel-level camera control and improvements to stabilization and motion understanding. The Extreme chip supports 8K video at 60 frames per second and 4K at 240 frames per second, along with Qualcomm’s Advanced Professional Video codec.
• Clearer calls: Both processors use AI to enhance vocals and reduce background noise. Qualcomm’s Voice Bubble technology is also designed to isolate a user’s voice during calls.

Why it matters

The announcement reflects a shift in the smartphone AI race. The question is no longer only whether a phone can execute a model, but whether local models can continuously understand a user and participate in everyday tasks. A sensing hub can collect relevant signals, an on-device model can process them quickly, and an agent can turn the result into an interaction or suggested action.

That architecture could reduce reliance on cloud services and keep some features available when connectivity is limited. However, the material does not provide concrete latency, battery, privacy, or benchmark data. The practical value of the new capabilities will therefore depend heavily on how phone makers integrate Qualcomm’s hardware into their operating systems and applications.

The 30B claim also needs context. Qualcomm’s Extreme chip is described as supporting a larger mixture-of-experts model than Apple’s 20B model announced at this year’s Worldwide Developers Conference. Yet parameter count alone does not determine user experience. The number of active parameters, memory bandwidth, model quality, and software optimization will all matter.

Motorola has announced the Motorola Signature 27, powered by the Snapdragon 8 Elite Extreme Gen 6, with general availability expected sometime this year. Qualcomm also says it is working on more than 40 AI devices. If those products deliver on the promise, the definition of an AI phone may increasingly depend on local agents, voice, and system-level personalization rather than on a standalone AI gadget.

Source: TechCrunch AI