Huawei details how the Kirin 9050 runs a 30-billion-parameter model on the device

Huawei detailed this week how the Kirin 9050 manages to run, entirely on the device and without an internet connection, an artificial intelligence model with 30 billion parameters. The explanation came in technical material released by the Chinese manufacturer and reported by the site TechNode, which also cites coverage from the Chinese site IT Home about the same chip.

The central point is the Da Vinci neural processing unit (NPU), the new version embedded in the Kirin 9050, which combines low-bit quantization with tighter coordination between memory and compute capacity. In practice, this is what makes it possible to fit a model of this size within the power and space constraints of a smartphone, something that would normally require a server GPU running in the cloud.

Why 30 billion parameters fit in a phone

The model cited by Huawei uses a mixture-of-experts (MoE) architecture, a technique in which the neural network is split into several specialized sub-models (the "experts") and only a fraction of them is activated for each query. According to Huawei, about 2 billion of the total 30 billion parameters are activated at a time, which drastically reduces the computational cost of each inference without giving up the model's total size.

That number, the 2 billion active parameters, had already been disclosed by Huawei itself in October, during a technical briefing about the chip. What the company detailed now is how the Da Vinci NPU sustains this architecture in practice: using low-bit quantization and coordinating memory and compute capacity to enable the model to run on the device, without Huawei having publicly detailed, in the disclosure reported by TechNode, the exact mechanism behind this coordination.

In short: the trick isn't shrinking the model, it's running only the piece of it that the task at hand actually needs.

Low-bit quantization does the rest of the work

In addition to partial activation via MoE, Huawei described the use of low-bit quantization to further reduce the model's memory footprint. Quantization is the technique of representing the neural network's weights with fewer bits than the training standard, trading some numerical precision for gains in speed and memory savings.

The combination of the two techniques, fewer active parameters at a time and each parameter taking up less space, is what Huawei points to as what makes it feasible to run a 30-billion-parameter model on mobile hardware. The company did not detail, in the disclosure reported by TechNode, the exact number of bits used in the quantization nor latency or battery consumption benchmarks for this workload.

The features Huawei already shows working

The same processor, according to Huawei, already supports end-user features that depend on this local AI capability:

• AI-assisted screen reading (AI screen reading), which interprets displayed content without needing to send data to a server;
• Eye-tracked page turning, which uses the camera to detect the user's eye movement and advance the content being read.

Both features depend on fast, constant inference, exactly the scenario in which running the model in the cloud would introduce noticeable latency and require a permanent internet connection, which underscores the reason Huawei is investing in keeping processing on the chip itself.

What this means for developers

For those building AI applications in Brazil, the Kirin 9050 case is a snapshot of a larger trend: large models no longer being synonymous with a cloud API. The MoE architecture combined with aggressive quantization is the same strategy other mobile chip makers have been pursuing to run language and computer vision models directly on the device, without the recurring cost of calls to external servers.

In practice, this opens the way for apps that work without internet, reduce exposure of sensitive user data (since nothing leaves the device), and don't depend on network latency to respond. For those developing embedded AI features, the point to keep in mind is that this kind of gain usually requires adapting the model to the MoE architecture and applying quantization specific to the target hardware; it's not something you get just by running a generic 30-billion-parameter model downloaded from the internet on any NPU.

What remains open

Huawei's disclosure, as reported by TechNode, does not include independent performance benchmarks, battery consumption during sustained inference, or a direct comparison with competing chips from other manufacturers that are also investing in on-device AI. There is also no detail, in the available sources, about which 30-billion-parameter model this is, whether it's a proprietary Huawei model or an adaptation of an open model, nor a timeline for broader availability of the screen reading and eye-tracking features for external developers to build on top of.

Source 1: TechNode (China) (https://technode.com/2026/11/11/huawei-details-how-kirin-9050-runs-a-30-billion-parameter-model-on-device/)

Huawei details how Kirin 9050 runs a 30-billion-parameter model on device · TechNode

Huawei on Wednesday detailed how the Kirin 9050 series runs a 30-billion-parameter AI model locally without an internet connection. The chip's new Da Vinci neural processing unit uses low-bit quantization and coordinates memory and compute to run the model on a phone.

The mixture-of-experts model activates about 2 billion parameters at a time. Huawei previously disclosed that figure at its October chip briefing and this week described how the same processor supports features such as AI screen reading and eye-tracked page turning. [IT Home, in Chinese] [October briefing, in Chinese]