Snapdragon 8 Elite Gen 6: Qualcomm Brings AI Agents Right Onto the Phone

At its Snapdragon Summit, Qualcomm unveiled two new flagship chips for Android smartphones: the Snapdragon 8 Elite Gen 6 and the more powerful 8 Elite Extreme Gen 6. Both are designed to run AI agents directly on the phone, without sending every request to the cloud. That is more than a marketing line: the Extreme version is supposed to run a language model with 30 billion parameters locally. The chips arrive, however, at a time when expensive memory is shrinking the smartphone market.

Key takeaways

• Qualcomm is introducing two flagship chips, the Snapdragon 8 Elite Gen 6 and the 8 Elite Extreme Gen 6, built on TSMC’s 2-nanometer process and reaching CPU clock speeds of up to 5 GHz for the first time.
• According to Qualcomm, the Extreme variant can run mixture-of-experts models with more than 30 billion parameters directly on the device.
• A low-power sensing hub runs small models of up to 200 million parameters continuously in the background, for example for local conversation transcripts.
• Xiaomi, OnePlus, Oppo, Honor and Motorola are among the brands announcing devices. At the same time, research firm Counterpoint expects smartphone shipments to fall 14 percent in 2026.

What the chips can do

At the core is Qualcomm’s in-house Oryon CPU. The two prime cores run at up to 5 GHz, the remaining cores at up to 4 GHz. Qualcomm calls it the first mobile CPU to hit 5 GHz. Compared with the previous generation, the company cites 13 percent more CPU performance and 44 percent more graphics performance for the Extreme variant. The Hexagon NPU, the block that handles AI workloads, is said to be 35 percent faster and up to 33 percent more power-efficient.

The most striking claim concerns model size. The Extreme version is supposed to run so-called mixture-of-experts models with more than 30 billion parameters locally; in these models, only a small part of the network is active at each step, which makes them attractive for devices with limited computing power. One enabler is 50 percent more shared memory for the NPU. For comparison, at its developer conference in June, Apple introduced a 20-billion-parameter mixture-of-experts model as the most capable of its new foundation models.

Less spectacular but possibly more important in daily use is the redesigned Sensing Hub. According to TechCrunch, this low-power companion processor can run small models of up to 200 million parameters. That is meant to let the phone transcribe conversations locally, tell speakers apart and learn from usage patterns to suggest useful automations. Qualcomm says a complete voice agent that listens and responds can even run on the chip. There are also new security features meant to control agents’ access to sensitive data.

From app phone to agent phone

Qualcomm CEO Cristiano Amon described the direction at the launch: the smartphone is shifting from a model built around navigating apps to one in which agents complete tasks for the user. Google executive Rick Osterloh said on stage that Gemini will better understand the user’s context, such as whether they are in a car, thanks to the chip’s processing power and sensors.

The case for doing this locally is obvious. An agent that learns habits, transcribes conversations and accesses messages, calendars and photos handles the most personal data a person has. If the model runs on the device, that data doesn’t have to leave the phone, responses don’t take a detour through the network, and much of it works offline. How eagerly people are waiting for such assistants is shown by the success of Meta’s agent Muse, which isn’t available in the EU yet and runs in the cloud.

The limits belong in the picture, too. A local model with 30 billion parameters is impressive, but it doesn’t play in the same league as the big data center models. The strongest open model alone, Xiaomi’s MiMo-V2.6-Pro, has just over a trillion parameters. A division of labor is more realistic: the phone handles everyday and private tasks itself and passes harder ones to the cloud. How well that works in practice depends less on the chip than on the software Google and the phone makers build on top of it.

The memory problem

The timing is tricky. Sharply higher memory prices are driving up the cost of finished devices, and according to CNBC the industry is in a worldwide contraction. Counterpoint Research expects smartphone shipments to fall 14 percent in 2026. That hits local AI directly, because language models need one thing above all: memory. Even heavily compressed, 30 billion parameters take up roughly 15 gigabytes by a rough calculation, more than the total memory in many of today’s smartphones. The memory shortage could therefore help decide how many buyers can actually use the new AI features to their full extent.

Phone makers are responding by focusing on expensive premium devices. Qualcomm is pursuing a similar strategy with two flagship chips: the Extreme version targets the most expensive models, while the regular Gen 6 variant is meant to bring many of the AI features to a broader high-end tier.

When the first devices arrive

Qualcomm lists Honor, iQOO, Motorola, OnePlus, Oppo, Redmi, RedMagic, Vivo and Xiaomi as partners. Motorola has already announced the Signature 27 as a first device with the Extreme chip, which TechCrunch says will be available sometime this year. When the new phones will reach stores in Germany and at what prices is still open.

Outlook

With this generation, the smartphone is turning from a screen for apps into a platform for personal agents that compute locally and can therefore handle private data without passing it on. Whether that catches on will be decided in two places Qualcomm doesn’t fully control: the software that turns raw computing power into useful helpers, and the price of memory, which will determine whether local AI remains a luxury feature or reaches the mass market.