PrismML helps Apple explore 27B-parameter on-device AI for iPhone

Apple is exploring ways to run significantly larger AI models directly on iPhones, according to a report revealing meetings with PrismML. The startup has managed to shrink Alibaba’s Qwen 3.6 model, with 27 billion parameters, to run fully on an iPhone 17 Pro, surpassing Apple’s own 20‑billion‑parameter AFM 3 Core Advanced model. Unlike Apple’s sparse architecture, PrismML’s approach allows all parameters to remain active simultaneously, enabling more powerful on-device processing. Running larger models locally could reduce Apple’s reliance on Private Cloud Compute, lowering operational costs and improving user privacy. This development suggests Apple may expand Apple Intelligence features that operate entirely on-device in future iOS versions.

Read the full story on MacRumors →