[Maui, United States = DigitalToday reporter Suk Dae-geon] Through data placement design, the size of on-device AI models previously bound by smartphone memory capacity has grown to 30 billion parameters. Qualcomm is extending the same principle of keeping data close to compute units to CPUs.

Chris Patrick (크리스 패트릭), Qualcomm executive vice president and head of its mobile handset business, said in a Snapdragon Summit 2026 keynote on Sept. 22 (local time) that "the more capable AI agents become, the more they need access to larger models with stronger reasoning and richer context." He said that for agents to handle tasks on a user's behalf, they must reason through multiple steps, find the required context and call tools, meaning model size and the length of context that can be handled at once need to grow together.

Qualcomm's flagship mobile platform, the Snapdragon 8 Elite 6th Gen unveiled at Snapdragon Summit 2026, runs a 30 billion-parameter AI model on a smartphone. It loads only the needed parts from flash storage and keeps frequently used data inside the NPU. Qualcomm said shared memory on the Hexagon NPU has expanded by 50%.

Qualcomm previously said in a pre-briefing that it was aware of the industry's memory constraints and was innovating by moving intelligence closer to the processor. The design shown at the summit indicates Qualcomm has increased memory inside the NPU. The Hexagon NPU stores more context and KV cache in the expanded shared memory. KV cache is temporary memory that stores what the model has already read so it does not recompute it.

Qualcomm explained that if this data remains inside the NPU, there are fewer trips to external DDR memory, lowering latency. The newly introduced element accelerator supports context windows of up to 32,000 tokens. Qualcomm said prefill performance, the stage when a model first reads an input, has improved by up to 80% in some models.

Outside the NPU, it uses a mixture-of-experts (MoE) architecture together with flash storage. MoE is a structure that activates only the parts relevant to a request among models divided into multiple specialist domains. Compared with dense models in which most of the network participates in every inference step, it reduces actual compute and memory bandwidth requirements. In each token-generation step, the model activates about 3 billion parameters out of the total, and it runs the model directly from flash by combining UFS 5.0 with its own intelligent flash memory management system.

Qualcomm showed a demonstration of an agent running on this model in the keynote. The agent read a summit invitation, identified the date, flight and venue, and planned a three-day schedule. For the first day, it arranged travel and check-in to match a welcome dinner; for the second day, it scheduled meetings during free time; and for the third day, it planned around the keynote and a booth reception. It said the MoE architecture activates different experts at the understanding, reasoning and execution stages.

With the previous architecture, it could not run MoE models with more than 30 billion parameters. With the new platform, it improved the NPU structure, added the element accelerator and expanded memory inside the NPU. It is interpreted as having chosen an approach that increases model size while dividing data across NPU memory, DRAM and flash. Qualcomm said this structure forms the foundation for implementing generative AI with fast responsiveness and privacy protection on smartphones.

The design of keeping data close to compute units has also been applied to the CPU. Qualcomm said it supports tasks that require fast data movement such as gaming, video editing and multitasking. Patrick said "the new Oryon CPU's FlexCache provides an L2 cache shared pool across all eight cores and dynamically allocates resources where needed," adding, "today's AI agents reason, find context, call tools and keep moving work across CPUs," and that FlexCache makes that switching faster.