Qualcomm's next Snapdragon flagship chip wants your AI agents to stay on-device

Qualcomm is set to reveal the next generation of flagship chips, including two Elite variants of the Snapdragon 8 Gen 6 (likely naming), on September 22 at the next Snapdragon Summit. While the names remain a mystery, Qualcomm has been dropping hints about the upcoming platforms one by one. After revealing that the upcoming chips would feature the first smartphone CPU to breach the 5GHz mark, and then sharing details about the upgraded GPU that would offer improved frame upscaling tech, Qualcomm is sharing cues about the upgrades coming to the neural processing unit, or NPU.

Today, Qualcomm is sharing details about the next generation of the Hexagon NPU, built to enable on-device multimodal AI applications, especially agents. Two key highlights of the upcoming NPU, according to Qualcomm, are the new “Element Accelerator” and “much larger shared memory.”

The Element Accelerator is a specialized component that now sits next to the scalar, vector, and tensor units. Qualcomm says it is designed to speed up transformer inference, “helping agents respond faster, reason more efficiently, and deliver richer experiences” without adding power cost. Without revealing the exact bandwidth, Qualcomm says the four accelerator units — tensor, vector, scalar, and the new Element — now have more shared memory to reduce dependence on the phone’s main RAM module while exchanging data from the KV-cache.

In simpler terms, that means the NPU can hold more context, making room for heavier workloads, especially agentic ones.

Qualcomm also says the new Hexagon NPU will support new AI architecture built specifically for Mixture-of-Experts (MoE) models. This means larger models can function on device by activating only a fraction of the parameters using a “routed” model. MoE routes queries through specialized expert models based on the input, reducing compute and memory bandwidth demands. It further states that the architecture will allow models with up to 30B parameters to run locally.

The chip giant also says the new NPU offers “50% higher pre-fill performance, faster decoding throughput, enhanced speculative decoding, and higher overall tokens per second.” These claims apply specifically to models with INT4 precision and should not be taken as a direct 50% increase in AI performance. Together, these enhancements can help models deliver faster responses and complete multi-step tasks more quickly.

The biggest advantage here is improved local processing, and Qualcomm says the upcoming Snapdragon Elite chips‘ improvements combined make agentic workflow, not just regular AI queries, “more efficient and responsive.”