Ahead of the Snapdragon Summit event, Qualcomm has detailed the next-generation Hexagon NPU for its upcoming flagship mobile platforms, the Snapdragon 8 Elite Gen 6 and the Snapdragon 8 Elite Extreme Gen 6.
The official Qualcomm blog post reveals that the next generation of Qualcomm Hexagon NPU is built to run agentic AI. This means that instead of only handling AI workloads, the upcoming SoCs will be able to run AI workloads across apps, understand operational context and perform multi-step reasoning and action.
To achieve this, Qualcomm says that the upcoming Snapdragon processors will feature a new NPU-based Element Accelerator, which is designed to handle the transformer workloads that power agentic and generative AI.
Element Accelerator will work alongside the Hexagon NPU's existing scalar, vector and matrix extensions, accelerating the specific operations Large Language Models (LLMs) lean on most.
Alongside that, Qualcomm has made the NPU's shared memory subsystem 50% larger, allowing more model state, activations and intermediate tensors to stay on-chip instead of being fetched from the allocated DDR memory.
Hexagon NPU Can Handle 30 Billion Parameter Models
Moving on, the Hexagon NPU is built around Mixture-of-Experts (MoE) models. MoE contains many specialised sub-networks and for each token it generates, it activates only the relevant ones.
Qualcomm's own example says that a 30-billion-parameter MoE model can keep tens of billions of parameters available while activating only around 3 billion per token generation step.
In this way, smartphones with the new Hexagon NPU will be able to access the capability of a much larger model while doing roughly a tenth of the work at any given moment. Qualcomm says that the brand is partnering with leading memory and model providers to build this architecture out.
Aside from that, the NPU supports a wide precision range, such as INT2, INT4, INT8, FP8 and FP16. This gives developers trade off performance, memory, model quality and power depending on the task. Smaller models will still run routine work regularly.
For INT4 models specifically, Qualcomm claims up to 50% higher prefill performance, faster decoding throughput, improved speculative decoding and more tokens per second overall. For those unaware, prefill is the stage where a model reads input before generating anything.
That said, Qualcomm has not named the chip on which the new Hexagon NPU will rest, referring only to it as "next-generation premium mobile platform." The Snapdragon 8 Elite Gen 6 series is the obvious candidate.
Qualcomm also outlined its Adreno GPU Neural Fusion system and 5 GHz for Oryon Core CPUs. The new Hexagon NPU details complete the trifecta of Snapdragon SoC architecture before the next-gen platforms are unveiled on September 22, 2026. So, stay tuned.
Related Articles
