Apple is reportedly in talks with a startup that has technology to cut the memory use of large artificial intelligence (AI) models to as little as one-fifteenth. If its technology validation succeeds, Apple's on-device strategy—processing Siri's AI computations inside the iPhone rather than in the cloud—is expected to strengthen.
On the 14th (local time), U.S. business network CNBC reported that PrismML, a spinoff from the California Institute of Technology, unveiled two compressed versions of Alibaba's open-source model "Qwen." The company said it reduced a roughly 54GB, 27-billion-parameter model to under 4GB so it can run even on iPhone 15 and later products.
PrismML developed a technology that simplifies the representation units of model weights to cut memory usage by 10 to 15 times. The company said this can boost processing speed by six to eight times and reduce power consumption by three to six times. Babak Hachibi, chief executive officer (CEO), said multiple corporations, including Apple, are evaluating the technology.
However, high compression ratios can come with performance degradation. CEO Hachibi also acknowledged that the ability to remember factual information or answer accurately could decline in part. Analysts said commercial rollout will depend on whether stability can be secured across numerous queries and diverse device environments.
Apple handles relatively simple AI tasks on-device and routes complex requests to Private Cloud Compute in a dual-track structure. If an ultra-lightweight model is applied, more functions can run offline, reducing response latency and server operating expense while also raising the level of privacy protection.
Some expect that as AI model slimming spreads, demand for high-bandwidth memory (HBM) for data centers could slow. Others counter that because compute demand does not disappear but shifts to smartphones and PCs, it will not lead to an overall consolidation of semiconductor demand. Battery drain from continuous operation is also cited as a challenge to solve.