Your phone is about to do a lot more thinking without ever pinging a server. On September 22, 2026, at its annual Snapdragon Summit, Qualcomm announced two new flagship chips, and the headline number is hard to ignore: the top-end model can run a 30-billion-parameter AI model entirely on the device.
That's not a rounding error. Running large models locally, without shipping your data to the cloud and waiting on a response, has been the holy grail for phone chipmakers for a couple of years now. Qualcomm just said its silicon can do it.
The two chips are the Snapdragon 8 Elite Gen 6 and the Snapdragon 8 Elite Extreme Gen 6. Both are built on TSMC's 2-nanometer manufacturing process, according to CNBC, and both are aimed squarely at premium phones. Motorola, Xiaomi, and ZTE are already lined up to use them, and Motorola used the event to unveil its new Signature 27 phone, powered by the Extreme chip, with general availability expected later this year.
The standout feature is the accelerator built into the Gen 6, designed to run AI models more efficiently, and the even beefier setup in the Extreme version, which Qualcomm says can handle a 30-billion-parameter mixture-of-experts (MoE) model on-device. If you're unfamiliar with MoE, the idea is that the model doesn't have to fire up all 30 billion parameters for every task. It only activates the specific "experts," or sub-networks, relevant to what you're asking it to do. That's what makes running something this large on a phone battery even plausible.
Both chips also come with a new sensing hub that can run smaller models, up to 200 million parameters, dedicated to things like a personal scribe that can tell speakers apart, memory that builds up based on how you use your phone, and a full voice-in, voice-out agent that doesn't need to leave the device. On the camera side, the chips add pixel-level control for more pro-style photography, better stabilization, and support for 8K 60fps and 4K 240fps slow-motion video on the Extreme version, along with a new Advanced Professional Video codec.
Qualcomm didn't shy away from making this a competitive story. CNBC reports the new chips are meant to go head-to-head with Apple's A20 Pro. And the numbers invite the comparison directly: Apple's third-generation foundation models, announced at WWDC in June, topped out with a 20-billion-parameter MoE model. Qualcomm's Extreme chip claims 30 billion. Bigger isn't automatically better in AI, capability depends a lot on training and tuning, not just parameter count, but it's a real gap and Qualcomm clearly wants it noticed.
The framing war matters too. Apple has been calling iPhones "intelligent personal hubs." Qualcomm CEO Cristiano Amon described the shift in almost identical spirit, saying the industry is "going into this transition from what is a very phone-centric model to now an agentic-centric model for new experiences." Two companies, same pitch: your phone isn't just an app launcher anymore, it's supposed to be the thing doing the work for you.
Amon also brought Google's Rick Osterloh on stage, pointing to Gemini's ability to understand context, like knowing you're in a car, thanks to the chip's sensors and processing power. That's a meaningful signal. Qualcomm isn't just building for its own software, it's building the plumbing other companies' AI assistants will run on.
The timing isn't random. CNBC notes the broader smartphone market is being squeezed by a worldwide contraction tied to skyrocketing memory costs, which have pushed up prices on finished devices. Counterpoint Research expects unit shipments to shrink 14% in 2026, with another 1% drop possibly in 2027. Squeezed on volume, phone makers are leaning harder into the premium tier, where buyers have more money to spend and are more willing to pay for AI bragging rights. A chip that can plausibly say "runs a 30B model locally" is exactly the kind of spec that justifies a higher price tag when the rest of the market is shrinking.
It's also worth remembering who's making this bet. Qualcomm built its business on foundational CDMA patents back in the 1980s and has spent decades as the default silicon supplier for Android flagships, alongside legal fights over patent licensing along the way. Betting the next generation of Snapdragon chips almost entirely on agentic AI is a continuation of a pattern, chase the next platform shift early, license the story to as many manufacturers as possible, and let the ecosystem build around you.
Practically, this is a big deal for anyone shopping for a premium Android phone next year. Motorola, Xiaomi, and ZTE devices built on these chips will market on-device AI capability as a headline feature, not a footnote. It matters for developers building agentic apps, since a sensing hub that runs speaker-differentiated transcription and task memory locally opens up new categories of always-on assistant features that don't need constant connectivity. And it matters for the ongoing Apple-versus-Android AI argument, which just got a very specific number attached to it.
Thirty billion parameters running locally is a genuinely impressive engineering claim, and if it holds up in real-world phones rather than just carefully staged demos, it's a legitimate leap for on-device AI. But announcements at a chip company's own summit are not the same as independent benchmarks. Qualcomm has an obvious incentive to frame this as beating Apple's 20-billion-parameter model, and parameter count alone doesn't tell you how good the model actually is at anything. The real test comes when the Signature 27 and its Snapdragon-powered siblings show up in reviewers' hands. Until then, this is a strong opening bid in a race that both Apple and Qualcomm clearly believe is the next decade of the smartphone business.