Qualcomm unveiled its new top smartphone chip on September 22, claiming it can run a 30-billion-parameter mixture-of-experts model entirely on a phone — a figure that is the vendor's own, with no released test code and no error bars from any outside lab S¹. The company launched two new mobile chips that day, betting that AI computation belongs in your pocket rather than a data centre S¹.
My read: This is the most aggressive on-device AI claim I've seen from a phone chip maker, and I don't buy the 30B figure until someone independent runs it. Mixture-of-experts models activate only a fraction of their parameters per token, which makes "30B" sound bigger than the memory and compute it demands at any given moment. Qualcomm has every reason to frame this in the most impressive terms. What I'd watch for is whether the chip holds that model in memory while a real app runs, rather than only in a controlled demo.
The architecture underneath is Qualcomm's Oryon CPU, which the company says is the first mobile processor to reach 5GHz P³. Francisco Cheng, writing on Qualcomm's own engineering blog in August, calls it "the world's fastest mobile CPU" P³. That is a vendor claim — no independent benchmark has been published, no test methodology disclosed, and the figure comes from Qualcomm's own marketing P³.
Mixture-of-experts, or MoE, is a model design where only a subset of the network fires on any given query, so a 30B model might engage just a few billion parameters at a time. That distinction matters. A 30B dense model on a phone would need to hold all 30 billion parameters in memory at once, something no current handset can manage. An MoE spreads the load, but Qualcomm has not said which model it ran, how much memory it consumed, or at what speed S¹.
Qualcomm's AI Research team has already shipped mobile video generation code
Qualcomm's AI research team has been working toward this. A public GitHub repository called Neodragon, maintained by contributors from Qualcomm AI Research, shows on-device video generation built on diffusion transformers P⁴. The repo carries a BSD 3-Clause licence and has 25 stars, small by open-source standards but a window into where Qualcomm's researchers are spending their time: generative AI running on phone silicon P⁴.
For handset engineers evaluating the new silicon, the pitch means weighing whether their thermal and battery envelopes can sustain on-device inference without throttling — Qualcomm has not disclosed power consumption figures S¹.
What this means for consumers depends on whether phone makers adopt the chips and whether developers build for them. Qualcomm has not named which manufacturers will use the new silicon, which devices will carry it, or when those phones will reach stores S¹. The company has also not disclosed pricing S¹.
As AI moves onto phones, the safety questions shift with it. When OpenAI launched its Australian Youth Safety Blueprint for teens earlier this week, the focus was cloud-based guardrails. On-device models change that equation, because a phone running its own model can bypass server-side filtering entirely.
The Neodragon repository remained public on GitHub under a BSD 3-Clause licence as of September 22 P⁴.
Sources: S1 — Qualcomm launches two new smartphone chips with emphasis on AI · S2 — Qualcomm launches two new smartphone chips with emphasis on AI - TechC · P3 — Qualcomm Oryon CPU reaches 5GHz with Flex Cache architecture · P4 — Qualcomm-AI-research/neodragon
Written from 4 sourced items, 2 of them primary.
