Qualcomm CEO Cristiano Amon announced that artificial intelligence companies want to run models with at least 100 billion parameters on phones in 2028. Amon stated in an interview on Alex Heath’s Sources program that this demand comes with the expectation of constantly working artificial intelligence. The statement doesn’t mean that the phones you can buy today support models of this scale. The executive describes the expectation from companies for future devices.
The conversation, recorded during the Snapdragon Summit in Hawaii, focuses on the role of phones in the age of artificial intelligence. Amon thinks agents that can use applications and understand personal context will create a new cycle of device refreshes. In his opinion, just as third-party applications have changed the phone experience, spies can trigger a similar transformation. However, this approach reflects the manager’s expectation rather than an actual result regarding sales.
Amon says that buyers who postponed switching phones due to increasing memory prices may also be affected by the new features. In addition, he explains that artificial intelligence companies are trying devices in the form of headphones, necklaces, badges and other forms. He sees smart glasses as the strongest candidate among wearable artificial intelligence products. He considers it necessary for the success of these devices that users trust the product, find it useful and easy to use.
The manager’s emphasis on 100 billion parameters does not indicate that the phone will run a single large model in the same way in all tasks. Qualcomm’s official description of the Hexagon NPU architecture highlights an approach that routes different tasks to specialized models. In this way, the company distinguishes between the total model size and the amount of active computation required in each operation.
Qualcomm mixes memory and expert networks for larger models
In its technical post dated September 10, the company gives an example of an expert mix model with 30 billion parameters. This model can enable approximately 3 billion directed parameters in each new token generation. Therefore, the total number of parameters does not mean that all parameters are included in the calculation together at each step. Qualcomm notes that expert selection and caching can reduce processing load.
The new Hexagon design uses the Element Accelerator unit for converter models and shared memory that grows by 50 percent. Keeping more intermediate data in memory aims to reduce accesses to external memory. In addition, the company supports different sensitivities between INT2 and FP16. These options allow developers to establish different balances between memory usage, model quality and energy consumption.
Qualcomm also reports that initial context processing performance can increase by up to 50 percent on INT4 models. However, this rate does not guarantee that all phones will provide the same speed gain in all artificial intelligence tasks. It is necessary to evaluate Amon’s request for 2028 and the company’s current technical examples separately. When choosing a phone today, working functions, supported applications and the actual features offered by the manufacturer continue to be decisive.