Apple is engaged in talks with PrismML, a stealth startup spun out of the California Institute of Technology, as it explores ways to move large-scale artificial intelligence workloads from remote servers back onto users’ iPhones.
The conversations follow a technical milestone from PrismML: the company compressed Alibaba’s 27-billion-parameter open-source Qwen 3.6 model so it can run locally on an iPhone 17 Pro. According to the startup’s reported results, the transformation shrank the model’s storage requirement from about 54 GB to less than 4 GB, while keeping every one of the model’s 27 billion parameters active at once and preserving benchmark performance.
Bringing models of that size onto a handset typically relies on a sparse architecture - a design that activates only portions of a model at any given moment to limit heat and power draw. PrismML’s approach departs from that pattern. The startup’s core technique uses ultra-dense 1-bit and ternary weight architectures, which it says reduce memory footprint as much as 14x and can run up to 8x faster.
Technical characteristics reported by PrismML:
- The Shrink - compressed the model from roughly 54 GB to under 4 GB.
- The Brainpower - preserved all 27 billion parameters active simultaneously without losing benchmark performance.
- The Core Tech - relied on 1-bit and ternary weight architectures to achieve large memory reductions and speedups.
For Apple, on-device execution of large models has several attractive outcomes: stronger user privacy because data need not leave the device, near-instant responses due to local processing, independence from cellular connectivity, and potentially large savings on cloud compute bills.
At its developer conference, Apple introduced an updated Siri architecture. Despite that change, a number of the more advanced features Apple announced still route data off-device to models hosted in the cloud, including Google’s Gemini models. By seeking companies such as PrismML, Apple appears to be pursuing a longer-term strategy to relocate heavy inference workloads back onto handsets.
That pivot would place Apple on a different trajectory from several large competitors. The article notes that firms such as Meta, Microsoft, and Amazon continue to invest heavily in large AI data centers, committing substantial capital to expand cloud-based capabilities. Apple’s method, if successful and deployable at scale, could reduce consumers’ dependence on those external data centers for everyday AI tasks.
What this means for markets and product strategy
Shifting heavyweight AI onto devices could change the cost dynamics for Apple’s services while reinforcing privacy and latency advantages that may strengthen device-level differentiation. For cloud operators and vendors of large-scale inference services, broader adoption of on-device models could affect future demand for centralized compute capacity.