Get the number right: 30 billion total, 2 billion active
On September 7, Huawei launched the Mate XT 2 tri-fold in Guangzhou and used the stage to unveil the Kirin 9050 Pro. The line that grabbed attention was this one: the industry's first on-device MoE omni-modal model, 30 billion total parameters, 2 billion activated.
Those two numbers only make sense together. MoE means mixture of experts. The pool holds 30 billion parameters, but each token only wakes a small slice of them, here about 2 billion. The part that actually does the work in any single inference step is one fifteenth of the total. That is not a knock. It is precisely why the model fits on a phone at all. But if you only remember "30 billion," you will expect far more capability than the device delivers.
What does it do in practice? The demonstration Huawei showed: with no network connection, running Pangu 30B-A2B locally, one spoken sentence is enough to organize a photo album. Working offline with data that never leaves the handset is the most concrete value on-device AI has.
How they fitted it in
The Kirin 9050 Pro uses an approach called logic folding. In plain terms: logic units used to be laid out flat on a single layer, like a one-story apartment. Now they are stacked in layers, like a duplex, with vertical channels between floors. Think of it as installing an elevator. Signals travel shorter distances, latency drops, and the same unit of time buys more work.
The formal name for the approach is the tau scaling law, first laid out systematically by He Tingbo, who leads Huawei's semiconductor business, at the IEEE ISCAS conference in May 2026. The core idea is to substitute time scaling for geometric scaling: no dependence on EUV lithography, stack logic units vertically on existing DUV equipment, and get close to advanced-node performance out of a mature process.
The measured numbers: transistor density went from roughly 155 million per square millimeter to about 238 million, a 55% increase in a single generation. He Tingbo said that gain equals the total progress from geometric shrinking over the previous three years.
The power figures are the counterintuitive part. Under real workloads at matched performance: NPU down 66%, GPU down 58%, CPU performance cores down 41%. Common sense says denser transistors run hotter. The explanation offered is that modern chips spend most of their energy moving data, not computing, and logic folding replaces long horizontal wiring with short vertical hops. At the same 29 TOPS of NPU throughput, the clock can drop 63% and the supply voltage falls from 0.85V to 0.55V, cutting power density by 73%.
The rest of the spec sheet: a nine-core CPU with simultaneous multithreading, up 24% single-core and 52% multi-core; the in-house Maliang GPU up 142% in rendering, with 4K multi-track export 40% faster. Overall system performance is up 42% over the previous generation. This is the first time since the Mate 40 in 2020, six years ago, that Huawei has devoted flagship launch time to a new Kirin chip.
Meanwhile, a second track is moving
On-device is not one company's story. ZTE's Nubia NaviX Ultra received network access approval from the Ministry of Industry and Information Technology. It ships with the Doubao phone assistant and is positioned as the world's first AI agent phone. Launch is set for September 16, with sales this month. The pitch is not raw model size. It is flipping the interaction model: instead of people adapting to a phone's menu system, the phone figures out what you are trying to get done.
Memory is moving in step. ChangXin Memory put LPDDR6 into volume production at 12,800 Mbps, with 10,667 Mbps LPDDR5X going into the NaviX Ultra. On-device models are bandwidth-hungry above all else.
A longer-shot path is also in play: diffusion language models are being pushed toward edge devices, with one team claiming a fivefold speedup for on-device agents. No verifiable public data yet.
Sober view: the line that was crossed, and the one that was not
What was crossed is whether it can run at all. Thirty billion total, two billion active, responding in real time on a phone: that was not possible a year ago. The three benefits are real. It works without a network. Data does not leave the device. There is no per-call token bill.
What was not crossed is how well it runs. On long-document comprehension, complex reasoning and multi-step tool use, phone-class models are a generation behind cloud flagships. Expecting local inference to match the cloud is not realistic in the near term. The likely shape is a division of labor: simple, private, high-frequency tasks on the device, heavy work in the cloud.
One point that gets missed: the stronger on-device capability gets, the more reason vendors have to keep data local, and that is not purely good news for users. Data that never leaves the phone means better privacy. It also means your on-device behavior becomes harder for third-party auditing tools to inspect. Benefits and costs often grow from the same stem.
What this means for you
If you are deciding whether an AI feature belongs on the device, check three conditions: does it need to work offline, is the data sensitive, is it called frequently. Hit two of the three and an on-device approach deserves a look. Hit none and stay in the cloud, where it is cheaper and better.
If you are just a phone buyer, the thing to watch right now is price. The Mate XT 2 starts at 19,999 yuan and tops out at 24,999, going on sale September 12. That is tri-fold flagship money, not for everyone. What will actually put on-device models in ordinary hands is the next wave of mid-range phones in the two-to-three-thousand-yuan band. The Mate 90 series is expected later this month on the same chip. If you can wait, wait for that one.