Thirty billion parameters. On a phone. Let that sink in for a second. Qualcomm is pushing hard to make your next Android flagship a legitimate AI powerhouse, and the Snapdragon 8 Elite Gen 6 might actually deliver on that promise. According to Android Headlines, the upcoming chip is being built around advanced on-device AI agents that can handle complex tasks without pinging a remote server every five seconds.
Qualcomm is set to officially pull back the curtain on September 22 at the Snapdragon Summit. The chip’s name isn’t confirmed yet, but “Snapdragon 8 Elite Gen 6” is what the leaks and online chatter are pointing to. Qualcomm’s naming has never exactly been straightforward, so take that with a grain of salt. What’s harder to dismiss is what the company has already teased ahead of the event: a CPU that breaks the 5GHz barrier and a new GPU with improved frame upscaling. But the real story here is the Hexagon NPU.
The updated Hexagon NPU is getting a new dedicated component called the Element Accelerator. This sits alongside the existing scalar, vector, and tensor units and is specifically designed to speed up transformer-based AI processing. That’s the kind of architecture powering today’s large language models, so having dedicated hardware for it on a mobile chip is a meaningful step forward. The goal is for AI agents to respond faster and reason more efficiently, without offloading the work to the cloud.
Qualcomm is also increasing the amount of shared memory available across those processing units. More shared memory means the phone can hold more information in play while an AI task is actively running. It’s a bit like giving the NPU a larger working desk instead of constantly shuffling papers in and out of a filing cabinet. The result, in theory, is smoother and more capable AI performance during complex, multi-step tasks.
The 30-billion-parameter Mixture-of-Experts model support is the headline number, and it’s worth putting in context. Most on-device AI today runs much smaller, more compressed models, because phones simply don’t have the memory or processing headroom for anything bigger. If Qualcomm can genuinely run MoE models at that scale locally, it would close a significant gap between what your phone’s AI can do versus what you’d normally need an internet connection and a data center to handle.
Here’s what this all adds up to for regular users. AI agents on Android phones could start handling things like managing your calendar, drafting responses, summarizing documents, and making decisions across apps, all without sending your data to a server somewhere. That’s better for privacy, faster in practice, and still functional when your connection is spotty.
Still, the full picture won’t be clear until we see actual benchmarks and real-world performance after launch. Qualcomm is good at setting expectations, and the Snapdragon Summit is always a well-produced show. But specs on a slide and chips in the wild are two different things. September 22 can’t come soon enough.