The new mobile processors can run 30 billion parameter AI models entirely on-device, signaling a major shift toward local inference and away from cloud dependence
Qualcomm just pulled back the curtain on its next generation of flagship mobile silicon, and the headline number is a big one. The Snapdragon 8 Elite Gen 6, set for a full unveiling at the Snapdragon Summit on September 22 in Maui, can run a 30 billion parameter mixture-of-experts AI model locally on a smartphone.
What’s actually under the hood #
The 30 billion parameter figure comes with an important asterisk that makes the whole thing work. The chip uses a Mixture-of-Experts (MoE) architecture, meaning the full model contains 30 billion parameters, but only about 3 billion are activated during any given inference step.
The key enabler is Qualcomm’s revamped Hexagon NPU, which now includes a dedicated “Element Accelerator” built specifically for transformer workloads. The NPU also gets 50% more shared memory compared to the previous generation, which allows context windows up to 32K tokens.
On the CPU side, the chip uses Qualcomm’s custom Oryon cores in a 2+3+3 configuration: two high-performance cores, three balanced cores, and three efficiency cores. The Adreno GPU gets an upgrade too, with new Matrix cores designed for AI-assisted graphics rendering.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Both the standard and Pro variants are expected to be built on TSMC’s cutting-edge 2nm process node. For the Pro version specifically, Qualcomm is adding extra Matrix ALU blocks to the GPU for what the company calls “AI Frame Fusion” capabilities.
The privacy play #
Beyond raw performance, Qualcomm is leaning hard into a narrative about on-device AI as a privacy advantage. The company frames its new chips as enabling “agentic AI” workflows, where an AI assistant can take multi-step actions on your behalf, without routing sensitive data through remote servers.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our