Qualcomm's New Snapdragon Chips Run 30B-Parameter AI Models On-Device

Qualcomm introduced two new smartphone processors this week, with the company's top-tier part capable of running a 30-billion-parameter mixture-of-experts model entirely on the device. The announcement positions on-device AI as the central selling point for the next generation of Android flagships, rather than an add-on feature. If the claim holds up in shipping hardware, it would mark a meaningful jump in what phones can do without a network connection.

The headline figure is the model size. A 30B mixture-of-experts network is far larger than the 7B and 13B models that current flagship phones typically handle, and MoE architectures activate only a fraction of their parameters per token, which keeps compute demands lower than the raw count suggests. Qualcomm is betting that combination — large sparse models plus dedicated NPU silicon — is the practical path to useful local assistants, summarizers and image tools.

Memory is the constraint that has kept big models off handsets. Running a 30B MoE locally requires the weights to fit in RAM alongside the operating system and apps, so the achievement depends as much on memory bandwidth and capacity as on raw TOPS. Qualcomm has not detailed the exact quantization or memory footprint used to hit the number, details that will determine whether the demo translates to everyday use.

The launch lands in a crowded field. Google has pushed Gemini Nano onto Pixel devices, Apple has built Apple Intelligence around on-device processing with cloud fallback, and MediaTek has marketed its own generative-AI phone chips. Qualcomm supplies silicon to most of the Android ecosystem, so its roadmap shapes what Samsung, Xiaomi, Oppo and others can offer within the next product cycle.

Running models locally also changes the privacy and cost math. Inference that stays on the phone avoids sending prompts to a server, cuts per-query cloud bills for device makers, and keeps working offline or on a weak connection. The tradeoff is that on-device models are typically smaller and less capable than their data-center counterparts, which is why most vendors still route hard queries to the cloud.

What to watch is benchmarking and battery life. Marketing numbers for model size say little about tokens per second, thermal throttling under sustained load, or how much the NPU drains a 5,000mAh battery during a long session. Independent tests on retail devices — not Qualcomm's reference hardware — will show whether a 30B local model is a genuine shift or a spec-sheet milestone.