TL;DR
Qualcomm's next-gen flagship mobile platform will feature a new Hexagon NPU with an Element Accelerator, large shared memory, Mixture-of-Experts (MoE) support and INT2–FP16 precision coverage, addressing continuous, multimodal, low-latency on-device agentic AI from the hardware architecture level.
Background: agentic AI changes mobile compute needs
As AI moves from one-shot Q&A to continuously sensing, cross-app agent scenarios, mainstream architectures are shifting from one giant model handling everything to systems of multiple specialized models scheduled by task, context and user intent. This demands not only faster AI inference but architecture designed for continuous, multimodal, low-latency operation.
On September 11, 2026, Cisco Cheng, Senior Director of Product Marketing at Qualcomm Technologies, published a blog describing the new Hexagon NPU for the next-gen Snapdragon flagship platform, with two key features: Element Accelerator and an expanded large shared-memory system.
Element Accelerator: dedicated acceleration for Transformer workloads
The Element Accelerator is designed for the Transformer workloads of generative and agentic AI, combining vector and scalar compute units: vector units handle high-throughput AI math, while scalar units support agent decision logic, routing and orchestration. Together with shared memory, they accelerate key model operations to improve agent response speed and inference efficiency while preserving mobile energy efficiency.
Large shared memory: easing the memory bottleneck
The Hexagon NPU expands large shared memory, keeping multi-model states, context and KV-cache closer to the accelerator. This reduces off-chip memory access and eases bandwidth pressure, letting agents stay responsive with longer context, more tool calls and concurrent tasks, while enabling longer context windows and faster token generation.
MoE and multi-precision support
The NPU is designed for next-generation model architectures. Qualcomm is working with memory and model vendors to bring Mixture-of-Experts (MoE) to on-device AI. For a 30B-parameter MoE model, the NPU needs to activate only ~3B routed parameters per generated token, cutting compute load and memory bandwidth.
For precision, the platform covers INT2, INT4, INT8, FP8 and FP16, letting developers trade off performance, memory, model quality and power. For INT4 models, it can deliver up to 50% higher prefill performance while accelerating decode throughput and speculative decoding to improve tokens-per-second.
What on-device agentic AI means
Agentic AI is defined by continuous operation and instant response, which requires compute to stay close to the user and on the device rather than in the cloud. Through the combination of accelerator units, shared memory, multi-precision support and model-loading mechanisms, the next-gen Hexagon NPU brings more agentic capability on-device, balancing responsiveness, privacy and power. The technology will reach consumers with the next-gen Snapdragon flagship platform.
Data sources & references
- Huanqiu Tech (September 11, 2026): Agentic AI reshapes mobile compute; Qualcomm details next-gen Hexagon NPU